@onescience/onecode-web 1.14.50-202609020932 → 1.14.50-202609041106

package/assets/catalog-Cz3JtNfO.js

@@ -1,273 +0,0 @@
-const a=[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球大气、区域降水、海洋涡旋"},{id:"inputData",label:"输入数据",placeholder:"例如：ERA5 场、雷达序列、归一化统计量、预训练权重",multiline:!0},{id:"goal",label:"任务目标",placeholder:"希望得到什么结果",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 多卡推理"}],e=[{id:"system",label:"研究对象 / 流场场景",placeholder:"例如：翼型、汽车外流、管道、Darcy"},{id:"inputData",label:"输入数据",placeholder:"例如：网格场、点云、图结构、CFD Benchmark 配置",multiline:!0},{id:"goal",label:"任务目标",placeholder:"希望得到什么结果",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / CFD Benchmark"}],t=[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：蛋白序列、复合物、配体口袋"},{id:"inputData",label:"输入数据",placeholder:"例如：FASTA、PDB、MSA、配体文件、模型权重",multiline:!0},{id:"goal",label:"任务目标",placeholder:"希望得到什么结果",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],l={id:"alphafold",name:"AlphaFold",category:"onemodel",domain:"bio",tags:["生信","结构预测","蛋白","MSA","RunModel"],blurb:"JAX/Haiku AlphaFold：MSA 特征经 RunModel 预测单体/多聚体结构与 pLDDT/PAE。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/alphafold",intro:{title:"30 秒看懂模型",subtitle:"用 DataPipeline 特征 + RunModel.predict 做单体/多聚体折叠，输出结构与置信度",suited:["单体折叠","多聚体推理","MSA 特征构建","置信度排序"],notFor:["核酸/配体复合物","无 MSA 管线","把 RunModel 当训练器"],prepare:["FASTA","MSA/template 或预计算特征","config 与 Haiku 参数"],obtain:["预测字典","pLDDT/PAE/pTM","ranking_confidence","可落盘 PDB"],steps:["确认 monomer/multimer","跑 DataPipeline","process_features","predict","置信度落盘"]},install:{banner:"已为你创建 AlphaFold 环境配置任务",prompt:["请为 AlphaFold 创建一个可以直接跑通的环境配置任务。","模型定位：JAX/Haiku AlphaFold 推理封装，公开入口 RunModel（process_features / predict）。","适合任务：MSA 特征构建、RunModel 单体折叠、Multimer 复合物推理、置信度排序落盘。","典型输入：FASTA / MSA·template 特征 / ConfigDict / Haiku params","典型输出：prediction dict / pLDDT·PAE·pTM / ranking / PDB","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience AlphaFold 入口（examples/biosciences/alphafold）与 onescience.flax_models.alphafold.model.model.RunModel 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：RunModel 单体折叠、Multimer 复合物推理、MSA 管线特征构建、置信度排序落盘。"].join(`
-`)},scenarios:[{id:"alphafold-monomer",title:"RunModel 单体折叠",summary:"DataPipeline 特征经 RunModel.predict 预测单体结构",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：单体激酶、抗体可变区"},{id:"inputData",label:"FASTA / 特征 / 权重",placeholder:"例如：query.fasta、msa 目录、params.npz",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：输出高置信单体 PDB 与 pLDDT",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"RunModel 单体推理",keyConfig:"multimer_mode=false、process_features、predict、random_seed",resources:"GPU/TPU；先 eval_shape 再完整推理",deliverables:"结构坐标、pLDDT、PAE、可序列化 PDB"},requiredData:["单体 FASTA 或 raw feature dict","与 config 匹配的 Haiku 参数树","DataPipeline 数据库或预计算 MSA"],expectedOutputs:"prediction dict、pLDDT/PAE 与 PDB 落盘路径",promptTemplate:["请基于 {{system}} 体系完成 AlphaFold「RunModel 单体折叠」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience AlphaFold（examples/biosciences/alphafold）执行：","1. 确认 config.model.global_config.multimer_mode=false，并加载匹配 Haiku 参数（勿把随机 init 当预训练结果）。","2. 用 DataPipeline 从 FASTA/MSA/template 生成 raw feature dict；FASTA 不能直接喂给 predict。","3. 实例化 RunModel(config, params)，调用 process_features(raw_features, random_seed)。","4. 可选 eval_shape(feat) 校验 JIT shape，再 predict(feat, random_seed)。","5. 输出结构字段与 pLDDT/PAE，并说明如何序列化为 PDB 与下一步质量评估。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"alphafold-multimer",title:"Multimer 复合物推理",summary:"多链配对 MSA 后预测蛋白复合物与 ipTM",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：抗体-抗原、异源二聚体"},{id:"inputData",label:"多链 FASTA / Multimer 特征",placeholder:"例如：chains.fasta、paired MSA、multimer params",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：评估界面 ipTM 与 PAE",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"AlphaFold Multimer 推理",keyConfig:"multimer_mode、MSA pairing、chain merge、ipTM",resources:"大显存 GPU；链数多时先短链冒烟",deliverables:"复合物结构、ipTM/pTM、ranking_confidence"},requiredData:["多链 FASTA 或 Multimer raw features","multimer 专用 checkpoint","链 ID 与配对 MSA 约定"],expectedOutputs:"复合物 PDB、ipTM/pTM 与 ranking 表",promptTemplate:["请基于 {{system}} 体系完成 AlphaFold「Multimer 复合物推理」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Multimer 流程执行：","1. 确认 multimer_mode=true，禁止与 monomer 参数/特征交叉混用。","2. 用 DataPipelineMultimer 构造多链特征、MSA pairing 与 chain merge。","3. RunModel.process_features → predict，读取 pTM/ipTM 与 ranking_confidence。","4. 检查界面 PAE 与链间接触，给出是否进入下游实验/对接的结论。","5. 若 shape 失败，先 eval_shape 定位字段，再回到管线重建特征。","如需安装依赖或提交远程作业，请先向我请求批准。"].join(`
-`)},{id:"alphafold-msa",title:"MSA 管线特征构建",summary:"从 FASTA 与数据库搜索产出 RunModel 可用 raw features",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：目标蛋白家族"},{id:"inputData",label:"FASTA / 数据库路径",placeholder:"例如：query.fasta、uniclust/bfd、template 截止",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出可直接 process_features 的特征包",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"构建 AlphaFold DataPipeline 特征",keyConfig:"JackHMMER/HHblits、template、residue_index、mask",resources:"CPU/大内存登录节点；数据库 I/O 密集",deliverables:"raw feature dict、字段检查报告、管线日志"},requiredData:["查询 FASTA","MSA/template 数据库与工具","monomer 或 multimer preset"],expectedOutputs:"raw feature dict、数据卡与下一步 RunModel 入口",promptTemplate:["请为 {{system}} 整理可供 AlphaFold RunModel 使用的 MSA/template 特征。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：","1. 选择 monomer DataPipeline 或 Multimer DataPipelineMultimer，并记录数据库版本。","2. 从 FASTA 跑序列搜索与 template hit，写出含序列/MSA/template/mask/residue_index 的 raw feature dict。","3. 核对字段完整性与 monomer/multimer 契约，形成简短数据卡。","4. 给出下一步可直接 RunModel.process_features → predict 的入口命令。"].join(`
-`)},{id:"alphafold-confidence",title:"置信度排序落盘",summary:"从 prediction dict 计算 ranking 并写出结构/置信度",fields:t,capabilitySummary:{current:"get_confidence_metrics 与结果落盘",keyConfig:"pLDDT、PAE、pTM/ipTM、ranking_confidence",resources:"本地 CPU 即可；大批量可并行写盘",deliverables:"ranking 表、pLDDT/PAE 摘要、PDB/CIF"},requiredData:["prediction_result 字典","multimer_mode 标志","目标输出格式"],expectedOutputs:"ranking 表、置信度 JSON 与结构文件",promptTemplate:["请基于 {{system}} 对 AlphaFold 预测结果做置信度排序与落盘。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请调用 get_confidence_metrics(prediction_result, multimer_mode)，按 ranking_confidence 排序，","写出结构文件与 pLDDT/PAE/pTM 摘要，并标注低置信区域与建议复核项（补充 MSA 或改用 AF3/Protenix）。","如字段缺失，先对照 preset 列出实际存在的置信度键，不要伪造指标。"].join(`
-`)}]},d={id:"alphafold3",name:"AlphaFold3",category:"onemodel",domain:"bio",tags:["生信","结构预测","复合物","扩散","配体"],blurb:"JAX AF3：JSON fold input 预测蛋白/核酸/配体复合物，输出扩散样本与 PAE/PDE。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/alphafold3",intro:{title:"30 秒看懂模型",subtitle:"AF3 JSON + data pipeline + diffusion sampling，面向多实体复合物结构与 ranking",suited:["多实体复合物","蛋白-配体","蛋白-核酸","扩散 ranking"],notFor:["仅需 AF2 单体快筛","Protenix 组件改造","无 JSON/CCD 输入"],prepare:["AF3 JSON","MSA/template/CCD","model_dir","recycle/sample 设置"],obtain:["结构文件","ranking","PAE/PDE/pLDDT","可选 embeddings"],steps:["校验 JSON 实体","跑 data pipeline","选择 attention","diffusion 推理","检查 ranking"]},install:{banner:"已为你创建 AlphaFold3 环境配置任务",prompt:["请为 AlphaFold3 创建一个可以直接跑通的环境配置任务。","模型定位：JAX/Haiku AF3-style 复合物预测，入口 Model + featurise_input / get_inference_result。","适合任务：AF3 JSON 复合物预测、蛋白-核酸-配体联合折叠、Diffusion sample ranking、MSA/CCD 管线联调。","典型输入：AF3 JSON / MSA·CCD·bonds / model_dir / recycle·diffusion samples","典型输出：结构文件 / ranking / PAE·PDE·pLDDT / confidence JSON","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience AlphaFold3 入口（examples/biosciences/alphafold3）与 onescience.flax_models.alphafold3.model.model.Model 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：AF3 JSON 复合物预测、蛋白-核酸-配体联合折叠、Diffusion sample 与 ranking、MSA/CCD 管线联调。"].join(`
-`)},scenarios:[{id:"alphafold3-json",title:"AF3 JSON 复合物预测",summary:"从 fold input JSON 跑 pipeline+inference 得到复合物结构",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：蛋白-配体复合物、binder-target"},{id:"inputData",label:"AF3 JSON / model_dir",placeholder:"例如：fold_input.json、model_dir、output_dir",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 Top ranked CIF 与 confidence",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"AF3 JAX 复合物推理",keyConfig:"run_data_pipeline、run_inference、flash_attention、num_recycles",resources:"大显存 GPU；兼容优先 attention=xla",deliverables:"结构文件、ranking、confidence JSON"},requiredData:["符合 schema 的 AF3 JSON","匹配版本的 model_dir/CCD","输出目录与注意力后端"],expectedOutputs:"CIF/PDB、ranking 与 PAE/PDE 摘要",promptTemplate:["请基于 {{system}} 完成 AlphaFold3「AF3 JSON 复合物预测」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 examples/biosciences/alphafold3 执行：","1. 校验 JSON 实体（蛋白/RNA/DNA/配体/离子/共价键）与 CCD/bonds 一致性。","2. 按资源设置 run_data_pipeline / run_inference；缺库时可关闭 pipeline 并挂载预计算 MSA/template。","3. 选择 flash_attention_implementation（兼容优先 xla），设置 recycle 与 diffusion sample 数。","4. 经 featurise_input → Model.__call__ → get_inference_result 产出结构与置信度。","5. 检查 ranking、pLDDT、PAE/PDE 与配体几何，并给出下游筛选建议。","如需安装依赖或提交远程作业，请先向我请求批准。"].join(`
-`)},{id:"alphafold3-multi",title:"蛋白-核酸-配体联合折叠",summary:"多分子类型同批预测并检查实体映射",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：核糖核蛋白+小分子"},{id:"inputData",label:"多实体 JSON / CCD",placeholder:"例如：protein+RNA+ligand JSON、CCD 组件",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证链映射与接触概率",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"多实体 AF3 联合推理",keyConfig:"token/atom 映射、CCD、bond、contact probability",resources:"大显存；长链降低 sample 或拆任务",deliverables:"多实体结构、实体映射表、接触概率"},requiredData:["含核酸/配体的 AF3 JSON","CCD 与化学常量版本","链/残基/配体 ID 约定"],expectedOutputs:"多实体结构与实体/链映射报告",promptTemplate:["请基于 {{system}} 做 AlphaFold3 蛋白-核酸-配体联合折叠。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认各实体类型被 AF3 支持，核对 atom/token 映射与共价键，","推理后重点检查核酸骨架、配体手性/碰撞与跨实体 PAE，输出实体映射与失败诊断。","不要把 Protenix feature dict 或 OpenFold batch 直接喂给 AF3 Model。"].join(`
-`)},{id:"alphafold3-rank",title:"Diffusion sample 与 ranking",summary:"多样本扩散采样后按置信度排序候选",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：已有 featurised BatchDict 的体系"},{id:"inputData",label:"BatchDict / sample 数",placeholder:"例如：features、num_diffusion_samples、rng_key",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：选出稳定复合物构象",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"AF3 扩散采样与 ranking",keyConfig:"diffusion samples、PAE/PDE、ranking confidence",resources:"GPU；显存紧时减少 samples、不存 pair embeddings",deliverables:"多样本结构、ranking 表、置信度曲线"},requiredData:["已 featurise 的 BatchDict","model params 与 rng_key","目标 sample 数"],expectedOutputs:"按 ranking 排序的结构与置信度报告",promptTemplate:["请基于 {{system}} 完成 AlphaFold3「Diffusion sample 与 ranking」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请设置 diffusion sample 数与 recycle，运行 Model.__call__，","用 get_inference_result 汇总各样本结构与标量置信度，按 ranking 输出 Top-N，","并说明低置信区域是否需补充 MSA/template 或用 Protenix 交叉验证。"].join(`
-`)},{id:"alphafold3-ccd",title:"MSA/CCD 管线联调",summary:"联调 data pipeline、CCD 与 featurisation 使 JSON 可推理",fields:t,capabilitySummary:{current:"AF3 data pipeline 与 CCD 联调",keyConfig:"DataPipeline.process、featurise_input、CCD/bonds",resources:"CPU 搜库 + GPU 冒烟；先小分子验证",deliverables:"可推理 BatchDict、管线日志、缺失项清单"},requiredData:["AF3 JSON","MSA/template 数据库或预计算","CCD/化学组件文件"],expectedOutputs:"BatchDict、缺失依赖清单与推理入口",promptTemplate:["请为 {{system}} 联调 AlphaFold3 MSA/CCD 管线。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：","1. 检查 JSON schema、CCD 组件与共价键定义是否一致。","2. 运行 DataPipeline.process，记录 MSA/template 覆盖；缺库则列出预计算替代方案。","3. 调用 featurise_input 生成 BatchDict，核对蛋白/核酸/配体 token 映射。","4. 给出最小 inference 冒烟命令与常见失败回退（attention→xla、减 sample）。"].join(`
-`)}]},p={id:"alphagenome",name:"AlphaGenome",category:"onemodel",domain:"bio",tags:["生信","基因组","调控","变异打分","DNA"],blurb:"JAX DNA 功能模型：对序列/区间/变异预测 tracks，并支持 interval 与 ISM 评分。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/alphagenome",intro:{title:"30 秒看懂模型",subtitle:"从本地 checkpoint 构建 AlphaGenomeModel，做 DNA 轨迹预测与变异/饱和突变评分",suited:["序列轨迹预测","区间功能打分","变异效应","ISM 扫描"],notFor:["蛋白结构任务","小分子对接","无参考基因组却要 interval API"],prepare:["checkpoint","organism FASTA/GTF","区间或变异表","OutputType"],obtain:["dna_output.Output","VariantOutput","scorer 分数","track metadata"],steps:["选 predict/score API","装配 OrganismSettings","create 模型","批量预测","下游排序"]},install:{banner:"已为你创建 AlphaGenome 环境配置任务",prompt:["请为 AlphaGenome 创建一个可以直接跑通的环境配置任务。","模型定位：JAX/Haiku DNA 功能预测，入口 create → AlphaGenomeModel.predict_* / score_*。","适合任务：DNA 序列轨迹预测、基因组区间功能打分、变异效应评分、饱和突变 ISM 扫描。","典型输入：checkpoint / organism FASTA·GTF / Interval·Variant / OutputType","典型输出：tracks / VariantOutput / scorer 分数 / metadata","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience AlphaGenome 入口（examples/biosciences/alphagenome）与 onescience.flax_models.alphagenome.model.dna_model.create 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：DNA 序列轨迹预测、基因组区间功能打分、变异效应评分、饱和突变 ISM 扫描。"].join(`
-`)},scenarios:[{id:"alphagenome-seq",title:"DNA 序列轨迹预测",summary:"对 DNA 字符串调用 predict_sequence 输出功能 tracks",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：启动子片段、增强子序列"},{id:"inputData",label:"DNA / checkpoint",placeholder:"例如：seq.fa、Orbax ckpt、OutputType 列表",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：预测表达/染色质相关 tracks",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"AlphaGenome predict_sequence",keyConfig:"checkpoint、OutputType、ModelSettings、device",resources:"GPU/TPU；~1Mbp 需大显存",deliverables:"dna_output.Output、track metadata"},requiredData:["DNA 字符串或 FASTA","匹配版本的 Orbax checkpoint","requested OutputType"],expectedOutputs:"轨迹张量/文件与 metadata",promptTemplate:["请基于 {{system}} 完成 AlphaGenome「DNA 序列轨迹预测」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用 create(checkpoint_path, ...) 构建 AlphaGenomeModel，调用 predict_sequence，","核对 OutputType 与 track shape/metadata；CPU 运行须显式传入 CPU device。","如需安装依赖或下载 checkpoint，请先向我请求批准。"].join(`
-`)},{id:"alphagenome-interval",title:"基因组区间功能打分",summary:"基于参考 FASTA 的 Interval 预测与 score_interval",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：hg38 某基因座"},{id:"inputData",label:"Interval / 参考基因组",placeholder:"例如：chr:start-end、FASTA、GTF、ontology",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：区间调控活性评分",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"predict_interval + score_interval",keyConfig:"FastaExtractor、OrganismSettings、ontology terms",resources:"GPU + 本地参考基因组索引",deliverables:"区间预测、scorer 分数、对齐报告"},requiredData:["0-based Interval","organism FASTA extractor","可选 GTF/PAS/splice 注释"],expectedOutputs:"区间 tracks 与 score_interval 结果",promptTemplate:["请基于 {{system}} 做 AlphaGenome 基因组区间功能打分。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请装配 OrganismSettings（FASTA/GTF 等），调用 predict_interval，再按需 score_interval；","缺 FASTA 时只能退回 predict_sequence 并明确限制；核对坐标约定与 organism/checkpoint 不可混用。"].join(`
-`)},{id:"alphagenome-variant",title:"变异效应评分",summary:"reference/alternate 对比预测并用 score_variant 打分",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：临床相关 SNP 集"},{id:"inputData",label:"Variant 表 / 参考",placeholder:"例如：VCF 或 variant CSV、FASTA、checkpoint",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：排序潜在调控变异",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"predict_variant + score_variant",keyConfig:"Interval+Variant、ref allele 与 FASTA 一致",resources:"GPU；大批量需批处理策略",deliverables:"VariantOutput、分数表、失败诊断"},requiredData:["Variant 与对齐 Interval","参考 FASTA（allele 一致）","checkpoint 与 organism"],expectedOutputs:"变异分数表与 ref/alt 预测摘要",promptTemplate:["请基于 {{system}} 完成 AlphaGenome「变异效应评分」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确保 reference allele 与 FASTA 一致，调用 predict_variant 与 score_variant，","输出可排序分数表；对大量 variants 使用稳定 interval 对齐与批处理。","scorer 缺注释时返回原始 prediction 并列出缺失依赖。"].join(`
-`)},{id:"alphagenome-ism",title:"饱和突变 ISM 扫描",summary:"对区间做 in-silico mutagenesis 评分扫描",fields:t,capabilitySummary:{current:"score_ism_variants",keyConfig:"ISM 变体集、splice/junction 依赖、batch 对齐",resources:"GPU；窗口过大时分块扫描",deliverables:"ISM 分数矩阵、敏感位点列表"},requiredData:["目标 Interval","ISM 变体生成策略","所需 splice/annotation 资源"],expectedOutputs:"ISM 分数热图/表与 Top 敏感位点",promptTemplate:["请基于 {{system}} 运行 AlphaGenome 饱和突变 ISM 扫描。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请调用 score_ism_variants（或等价 ISM 流程），保留 scorer 名称与参数，","输出位点级分数并标出高影响碱基；窗口过大时分块并合并结果。"].join(`
-`)}]},m={id:"diffdock",name:"DiffDock",category:"onemodel",domain:"bio",tags:["生信","分子对接","扩散","蛋白-配体","构象"],blurb:"蛋白-配体扩散对接：采样平移/旋转/扭转 pose，可用 confidence 模型重排序。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/diffdock",intro:{title:"30 秒看懂模型",subtitle:"给定受体 PDB 与配体 SDF/SMILES，扩散采样对接构象并用 confidence 重排序",suited:["单复合物对接","构象重排序","配体库批量 docking","口袋裁剪冒烟"],notFor:["新分子生成","亲和力回归主模型","蛋白骨架设计"],prepare:["受体 PDB","配体 SDF/MOL2/SMILES","score/confidence checkpoint","采样步数"],obtain:["候选 pose","score/confidence 表","日志","失败诊断"],steps:["清洗受体/配体","建 complex graph","加载 score 模型","扩散采样","confidence 排序"]},install:{banner:"已为你创建 DiffDock 环境配置任务",prompt:["请为 DiffDock 创建一个可以直接跑通的环境配置任务。","模型定位：E(3) 等变扩散对接，入口 build_score_model / load_score_model + 采样脚本。","适合任务：蛋白-配体扩散对接、Confidence 构象重排序、配体库批量 docking、口袋裁剪对接冒烟。","典型输入：受体 PDB / 配体 SDF·SMILES / score·confidence ckpt / samples_per_complex","典型输出：候选 pose / confidence 排序表 / 日志","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience DiffDock 入口（examples/biosciences/diffdock）与 onescience.models.diffdock（build_score_model / CGModel） 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：蛋白-配体扩散对接、Confidence 构象重排序、配体库批量 docking、口袋裁剪对接冒烟。"].join(`
-`)},scenarios:[{id:"diffdock-dock",title:"蛋白-配体扩散对接",summary:"对单个蛋白-配体复合物扩散采样对接 pose",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：激酶口袋 + 抑制剂"},{id:"inputData",label:"PDB / 配体 / checkpoint",placeholder:"例如：receptor.pdb、ligand.sdf、score model_dir",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 Top pose 并可视化检查",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"DiffDock 扩散对接采样",keyConfig:"samples_per_complex、inference_steps、batch_size、温度",resources:"GPU；大蛋白先裁剪口袋",deliverables:"候选 pose、score 表、采样日志"},requiredData:["受体 PDB（链 ID 清晰）","配体 SDF/MOL2/SMILES","score 模型目录与 checkpoint"],expectedOutputs:"pose 文件、排序表与失败诊断",promptTemplate:["请基于 {{system}} 完成 DiffDock「蛋白-配体扩散对接」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 examples/biosciences/diffdock 执行：","1. 清洗受体 PDB/配体键序与可旋转键；必要时裁剪结合口袋。","2. 经 biology_diffdock_dataset 构造 complex graph，确认 ESM 嵌入需求。","3. 加载 score 模型（model_parameters.yml + checkpoint），设置 samples_per_complex / inference_steps。","4. 运行采样并检查 pose 数量、碰撞与日志错误。","5. 将候选交给可视化或 GenScore 复筛；显存不足时降采样数或裁剪口袋。","如需安装依赖或提交远程作业，请先向我请求批准。"].join(`
-`)},{id:"diffdock-confidence",title:"Confidence 构象重排序",summary:"用 confidence 模型对 DiffDock pose 重排序",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：已采样 pose 的靶点-配体对"},{id:"inputData",label:"pose / confidence ckpt",placeholder:"例如：poses/、confidence model_dir、score 兼容类型",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：按 confidence 选出可靠构象",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"DiffDock confidence 重排序",keyConfig:"confidence checkpoint、与 score 模型类型兼容、温度",resources:"GPU 小批量即可；失败可关闭 confidence",deliverables:"重排序表、Top pose、兼容性检查"},requiredData:["已生成候选 pose","confidence 模型目录与 checkpoint","score 模型类型信息"],expectedOutputs:"confidence 排序表与推荐 pose",promptTemplate:["请基于 {{system}} 对 DiffDock 候选做 Confidence 构象重排序。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请核对 confidence 与 score 模型类型兼容后加载 checkpoint，","对候选 pose 打分并排序；若加载失败，先关闭 confidence 仅保留 score 采样结果并给出修复清单。"].join(`
-`)},{id:"diffdock-batch",title:"配体库批量 docking",summary:"对配体库 CSV/目录批量跑 DiffDock 对接",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：虚拟筛选苗头库"},{id:"inputData",label:"受体 / 配体库 CSV",placeholder:"例如：receptor.pdb、ligands.csv（SMILES/路径字段）",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：批量产出可排序 pose 清单",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"DiffDock 批量对接",keyConfig:"CSV 字段完整性、batch_size、输出目录映射",resources:"多卡或队列；先小子集冒烟",deliverables:"批量 pose、汇总表、失败配体列表"},requiredData:["固定受体 PDB","字段完整的配体库 CSV","score checkpoint 与输出根目录"],expectedOutputs:"批量结果表、失败清单与重试建议",promptTemplate:["请基于 {{system}} 完成 DiffDock 配体库批量 docking。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请先校验 CSV 字段齐全，用小子集冒烟确认图构建与采样，","再全量运行并汇总成功/失败配体；记录与上游蛋白结构 ID 的映射供下游打分。"].join(`
-`)},{id:"diffdock-pocket",title:"口袋裁剪对接冒烟",summary:"裁剪结合口袋后做短采样验证管线可跑通",fields:t,capabilitySummary:{current:"口袋裁剪 DiffDock 冒烟",keyConfig:"口袋残基、samples 小数、batch_size=1",resources:"单卡冒烟；通过后再放大采样",deliverables:"裁剪 PDB、冒烟 pose、资源结论"},requiredData:["全长或域 PDB","口袋定义（残基/半径）","score 模型可用性"],expectedOutputs:"裁剪结构、冒烟日志与正式采样参数建议",promptTemplate:["请为 {{system}} 做 DiffDock 口袋裁剪对接冒烟。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按口袋残基裁剪受体，用极少 samples_per_complex 验证建图与前向，","确认无显存错误后再给出正式批量对接参数；PDB 解析失败时清理链 ID/异质原子。"].join(`
-`)}]},c={id:"esm",name:"ESM",category:"onemodel",domain:"bio",tags:["生信","蛋白语言模型","表征","ESMFold","突变"],blurb:"蛋白语言模型族：ESM2 表征、ESMFold 快折、突变打分与 GVP 逆折叠对照。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/esm",intro:{title:"30 秒看懂模型",subtitle:"按任务选择 ESM2/ESMFold/ESM1v/GVP，完成表征、快折、突变评分或逆折叠",suited:["序列表征","ESMFold 快折","零样本突变打分","逆折叠对照"],notFor:["DNA/RNA 基因组主任务","默认替代 ProteinMPNN","超长无分段序列"],prepare:["蛋白 FASTA","模型权重","突变表或 backbone PDB","token batch 设置"],obtain:["embedding .pt","contact/PDB","突变分数 CSV","逆折叠 FASTA"],steps:["选定模式","检查长度与权重","API/CLI 推理","解析产物","下游复核"]},install:{banner:"已为你创建 ESM 环境配置任务",prompt:["请为 ESM 创建一个可以直接跑通的环境配置任务。","模型定位：ESM 蛋白语言模型族，入口 onescience.models.esm 预训练加载与 ESMFold/打分 API。","适合任务：ESM2 序列表征提取、ESMFold 快速折叠、零样本突变打分、GVP 逆折叠对照。","典型输入：蛋白 FASTA / ESM 权重 / 突变 CSV 或 backbone PDB / token batch","典型输出：embedding .pt / ESMFold PDB / 突变分数 CSV / 逆折叠 FASTA","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience ESM 入口（examples/biosciences/esm）与 onescience.models.esm 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：ESM2 序列表征提取、ESMFold 快速折叠、零样本突变打分、GVP 逆折叠对照。"].join(`
-`)},scenarios:[{id:"esm-embed",title:"ESM2 序列表征提取",summary:"提取 per-token/mean embedding 与可选 contact map",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：酶超家族序列集"},{id:"inputData",label:"FASTA / 模型名",placeholder:"例如：proteins.fasta、esm2_t33、repr layers",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出下游 ML 可用 .pt 特征",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ESM2 表征提取",keyConfig:"repr layer、include、max-tokens-per-batch、模型名",resources:"GPU；超长序列降 batch 或分段",deliverables:"embedding .pt、可选 contact、日志"},requiredData:["蛋白 FASTA","ESM2 权重或模型名","目标层与 include 项"],expectedOutputs:"embedding 文件与层配置记录",promptTemplate:["请基于 {{system}} 完成 ESM「ESM2 序列表征提取」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用 esm_sequence_batch_converter 与预训练加载提取指定层表征，","控制 max-tokens-per-batch；输出 .pt 并记录模型/层版本，供聚类或 SimpleFold 特征支持。","如需下载权重，请先向我请求批准。"].join(`
-`)},{id:"esm-fold",title:"ESMFold 快速折叠",summary:"从蛋白 FASTA 快速预测结构作初筛",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：设计序列初筛"},{id:"inputData",label:"FASTA / ESMFold 权重",placeholder:"例如：designs.fasta、esmfold 权重、chunk-size",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：快速 PDB 供 OpenFold/Protenix 复核",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ESMFold 结构预测",keyConfig:"recycle、chunk-size、CPU offload、batch",resources:"GPU；OOM 时 offload/增大 chunk",deliverables:"PDB、置信度摘要、失败序列"},requiredData:["蛋白 FASTA","ESMFold 权重","输出目录"],expectedOutputs:"ESMFold PDB 与初筛结论",promptTemplate:["请基于 {{system}} 运行 ESMFold 快速折叠。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请加载 ESMFold 权重对 FASTA 推理，设置 chunk-size/recycle；","结果仅作初筛，高价值候选建议 OpenFold/Protenix 复核；OOM 时启用 CPU offload 或改 SimpleFold。"].join(`
-`)},{id:"esm-variant",title:"零样本突变打分",summary:"用 ESM/ESM1v 对突变表做零样本效应评分",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：稳定性/功能突变扫描"},{id:"inputData",label:"野生型 / 突变 CSV",placeholder:"例如：wt.fasta、mutations.csv、esm1v 权重",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：排序优先实验突变",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ESM 零样本突变打分",keyConfig:"scoring strategy、offset、模型变体",resources:"GPU/CPU；大批量表格可分片",deliverables:"突变分数 CSV、排序表"},requiredData:["野生型序列","突变表（列与 offset 正确）","ESM1v/ESM2 权重"],expectedOutputs:"突变分数 CSV 与 Top 候选",promptTemplate:["请基于 {{system}} 完成 ESM 零样本突变打分。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请校验突变列与氨基酸字母表，按选定 scoring strategy 打分并输出 CSV；","高分突变可进入结构预测或实验排序，格式错误时先清洗再跑。"].join(`
-`)},{id:"esm-invfold",title:"GVP 逆折叠对照",summary:"对 backbone 做 ESM/GVP 逆折叠并与 ProteinMPNN 对照",fields:t,capabilitySummary:{current:"ESM/GVP inverse folding",keyConfig:"链 ID、采样温度、与 ProteinMPNN 对照协议",resources:"GPU；普通设计默认仍优先 ProteinMPNN",deliverables:"逆折叠 FASTA、对照表、日志"},requiredData:["backbone PDB/CIF","GVP/ESM inverse folding 权重","设计链与固定位点"],expectedOutputs:"逆折叠序列与 ProteinMPNN 对照摘要",promptTemplate:["请基于 {{system}} 做 ESM/GVP 逆折叠对照。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请明确这是对照路线而非默认 backbone→序列主路径；","从 PDB 采样/打分序列，与 ProteinMPNN 结果对照，并规划 SimpleFold/OpenFold 验证。"].join(`
-`)}]},u={id:"evo2",name:"Evo2",category:"onemodel",domain:"bio",tags:["生信","基因组","Mamba","长序列","DNA语言模型"],blurb:"基因组 Mamba 语言模型：对 token 化 DNA 做 logits、生成或带 loss_mask 训练。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/evo2",intro:{title:"30 秒看懂模型",subtitle:"基于 MambaModel.forward / mamba_forward_step 构建基因组 token 训练或推理",suited:["token 推理","长序列生成","FASTA 分词训练","并行 checkpoint 冒烟"],notFor:["未分词原始 FASTA 直喂","蛋白折叠主任务","忽略并行显存限制"],prepare:["Mamba 配置与 ckpt","tokenizer/vocab","input_ids/labels","并行设置"],obtain:["logits/hidden","训练 loss","生成序列","契约记录"],steps:["对齐 ckpt 与 tokenizer","构造 tokens","forward/生成","校验 shape","保存"]},install:{banner:"已为你创建 Evo2 环境配置任务",prompt:["请为 Evo2 创建一个可以直接跑通的环境配置任务。","模型定位：基因组 Mamba 语言模型，入口 MambaModel.forward 与 mamba_forward_step。","适合任务：基因组 token 推理、Mamba 长序列生成、FASTA 分词训练前向、并行 checkpoint 冒烟。","典型输入：Mamba config·ckpt / tokenizer / input_ids·position_ids·labels·loss_mask","典型输出：logits / loss / 生成序列 / 并行契约","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience Evo2 入口（examples/biosciences/evo2）与 onescience.models.evo2.MambaModel 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：基因组 token 推理、Mamba 长序列生成、FASTA 分词训练前向、并行 checkpoint 冒烟。"].join(`
-`)},scenarios:[{id:"evo2-infer",title:"基因组 token 推理",summary:"对 token 化 DNA 调用 forward 得到 logits/hidden",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：微生物基因组片段"},{id:"inputData",label:"tokens / checkpoint",placeholder:"例如：input_ids、position_ids、evo2 ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：计算序列概率或嵌入",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Evo2/Mamba 推理",keyConfig:"input_ids、position_ids、inference context、并行布局",resources:"GPU/多卡；长度受并行与显存限制",deliverables:"logits/hidden、shape 校验"},requiredData:["与 ckpt 匹配的 tokenizer","input_ids 与 position_ids","模型配置"],expectedOutputs:"logits 或 embedding 与配置记录",promptTemplate:["请基于 {{system}} 完成 Evo2「基因组 token 推理」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请从 checkpoint 对齐配置/vocab/并行，用源码 tokenizer 构造张量，","无 labels 时调用 MambaModel.forward；decoder_input 非空时勿重复 embedding。","如需安装 NeMo/Megatron 依赖，请先向我请求批准。"].join(`
-`)},{id:"evo2-generate",title:"Mamba 长序列生成",summary:"用 inference wrapper 做基因组序列续写/生成",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：启动子下游续写"},{id:"inputData",label:"prompt tokens / 长度",placeholder:"例如：prompt FASTA→tokens、max length、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成候选 DNA 片段",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Evo2 序列生成",keyConfig:"get_inference_wrapper、context 长度、采样参数",resources:"多卡并行；不可只加长 wrapper 忽略显存",deliverables:"生成 token/FASTA、长度统计"},requiredData:["prompt 序列与 tokenizer","匹配 checkpoint","目标生成长度"],expectedOutputs:"生成序列与资源使用说明",promptTemplate:["请基于 {{system}} 做 Evo2/Mamba 长序列生成。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请配置 inference wrapper 与并行，在显存允许范围内生成，","输出解码序列并做 tokenizer round-trip 抽查；OOM 时缩短长度或减小 micro batch。"].join(`
-`)},{id:"evo2-train",title:"FASTA 分词训练前向",summary:"从 FASTA dataset 构造四字段 batch 跑 mamba_forward_step",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：物种基因组训练集"},{id:"inputData",label:"FASTA / tokenizer",placeholder:"例如：train.fasta、vocab、labels/loss_mask 约定",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证训练 loss 正常下降冒烟",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Evo2 训练前向",keyConfig:"tokens、position_ids、labels、loss_mask、mamba_forward_step",resources:"多卡；先单步 loss 冒烟",deliverables:"loss、字段契约、冒烟日志"},requiredData:["FASTA 与 tokenizer","四字段 batch 映射","训练配置与 ckpt 规格"],expectedOutputs:"loss 曲线入口与 batch 契约说明",promptTemplate:["请基于 {{system}} 完成 Evo2 FASTA 分词训练前向。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用源码 FASTA dataset/tokenizer 构造 tokens/position_ids/labels/loss_mask，","复用 mamba_forward_step；校验 logits 与 labels shape 对齐，loss 异常时先查 mask 而非改输出层。"].join(`
-`)},{id:"evo2-smoke",title:"并行 checkpoint 冒烟",summary:"验证 checkpoint、并行布局与最小前向一致",fields:t,capabilitySummary:{current:"Evo2 checkpoint 冒烟",keyConfig:"并行布局、vocab、短序列 forward",resources:"与 ckpt 原生并行一致；先最短序列",deliverables:"加载报告、forward 成功日志"},requiredData:["checkpoint 与原生并行配置","匹配 tokenizer","短测试序列"],expectedOutputs:"冒烟通过报告与正式训练/推理入口",promptTemplate:["请为 {{system}} 做 Evo2 并行 checkpoint 冒烟。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 checkpoint 原生并行加载，跑最短序列 forward；","依赖版本不匹配时对齐 NeMo/Megatron/BioNeMo，不要改层名绕过加载错误。"].join(`
-`)}]},_={id:"genscore",name:"GenScore",category:"onemodel",domain:"bio",tags:["生信","亲和打分","图模型","虚拟筛选","蛋白-配体"],blurb:"蛋白-配体图打分：双塔编码后输出混合密度参数，经 scoring 得亲和分数。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/genscore",intro:{title:"30 秒看懂模型",subtitle:"用 GraphTransformer/GatedGCN 双塔与 GenScore.forward/scoring 做结合打分",suited:["亲和打分","双塔评估","原子贡献分析","虚拟筛选排序"],notFor:["生成对接 pose","新分子扩散生成","把七元组当最终标量"],prepare:["配体/口袋图","encoder 配置","checkpoint","配体 ID 映射"],obtain:["混合密度七元组","score/贡献","排序表","训练 loss"],steps:["构图","建双塔 encoder","加载 GenScore","scoring 后处理","排序"]},install:{banner:"已为你创建 GenScore 环境配置任务",prompt:["请为 GenScore 创建一个可以直接跑通的环境配置任务。","模型定位：蛋白-配体图打分，入口 GenScore.forward 与高层 scoring（非直接标量）。","适合任务：蛋白-配体亲和打分、双塔图编码器评估、原子贡献可解释分析、虚拟筛选候选排序。","典型输入：配体/靶点 PyG 图 / encoder 类型 / GenScore ckpt / dist_threhold","典型输出：混合密度参数 / score / 贡献 / 排序表","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience GenScore 入口（examples/biosciences/genscore）与 onescience.models.genscore.GenScore / scoring 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：蛋白-配体亲和打分、双塔图编码器评估、原子贡献可解释分析、虚拟筛选候选排序。"].join(`
-`)},scenarios:[{id:"genscore-score",title:"蛋白-配体亲和打分",summary:"对蛋白口袋-配体对计算最终亲和 score",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：对接后候选复合物"},{id:"inputData",label:"结构 / 图 / ckpt",placeholder:"例如：complex PDB、ligand SDF、genscore ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：输出可排序亲和分数",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"GenScore 高层 scoring",keyConfig:"in_channels、hidden_dim、n_gaussians、dist_threhold",resources:"GPU；批量按配体 ID 对齐",deliverables:"score 表、配体 ID 对齐检查"},requiredData:["配体与靶点图或可构图结构","匹配 checkpoint","配体 ID 列表"],expectedOutputs:"亲和分数表与对齐报告",promptTemplate:["请基于 {{system}} 完成 GenScore「蛋白-配体亲和打分」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请经 biology_genscore_dataset 构图，构造同输出通道 ligand/target encoder 与 GenScore，","加载 state_dict 后用 scoring 或 run_an_eval_epoch 得到最终分数（forward 七元组需后处理）。","核对 score 数量与配体 ID 一致。"].join(`
-`)},{id:"genscore-encoder",title:"双塔图编码器评估",summary:"评估 GraphTransformer/GatedGCN 双塔与特征维度一致性",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：更换 encoder 的打分实验"},{id:"inputData",label:"图 batch / encoder 配置",placeholder:"例如：PyG batches、GraphTransformer 参数、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：确认 encoder 输出契约可训练",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"GenScore 双塔评估",keyConfig:"GraphTransformer 或 GatedGCN、x/pos/batch、edge_index",resources:"GPU；先单 batch 前向",deliverables:"七元组、维度报告、loss 冒烟"},requiredData:["配体/靶点图 batches","encoder 构造参数","可选 checkpoint"],expectedOutputs:"维度一致性报告与最小 loss 结果",promptTemplate:["请基于 {{system}} 评估 GenScore 双塔图编码器。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用源码 _build_encoder 等价参数构造两侧 encoder，检查 x/pos/batch（配体含 edge_index），","跑 GenScore.forward；checkpoint 失败时重建准确 dims，勿在 forward 内广播不同样本。"].join(`
-`)},{id:"genscore-contrib",title:"原子贡献可解释分析",summary:"从打分结果提取原子/残基贡献用于解释",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：苗头化合物可解释性"},{id:"inputData",label:"score 输出 / 结构",placeholder:"例如：scoring 贡献结果、口袋残基编号",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：标出关键接触原子",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"GenScore 贡献分析",keyConfig:"贡献字段、原子/残基映射、可视化导出",resources:"本地即可；大批量可并行",deliverables:"贡献表、关键残基列表、报告"},requiredData:["已完成的 scoring 输出","原子/残基索引映射","结构文件"],expectedOutputs:"贡献分析报告与关键位点列表",promptTemplate:["请基于 {{system}} 做 GenScore 原子贡献可解释分析。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请从 scoring 贡献结果映射到残基/原子，输出 Top 贡献位点，","并说明如何用于报告或突变设计；缺少贡献字段时回到评估后处理而非改七元组。"].join(`
-`)},{id:"genscore-vs",title:"虚拟筛选候选排序",summary:"对接或生成后的大配体列表用 GenScore 复筛排序",fields:t,capabilitySummary:{current:"GenScore 虚拟筛选排序",keyConfig:"批量构图、score 阈值、Top-N",resources:"GPU 队列；先小库校准分数分布",deliverables:"排序表、Top-N 结构、阈值说明"},requiredData:["配体库与口袋结构","GenScore checkpoint","排序阈值/Top-N"],expectedOutputs:"虚拟筛选排序表与 Top-N 交付物",promptTemplate:["请基于 {{system}} 用 GenScore 做虚拟筛选候选排序。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请批量构图打分，按 score 排序并给出阈值与 Top-N；","保留与上游 docking/生成 ID 的映射，供 MD 或人工检查。"].join(`
-`)}]},h={id:"laproteina",name:"LaProteina",category:"onemodel",domain:"bio",tags:["生信","蛋白生成","flow-matching","骨架","motif"],blurb:"Product-space flow matching 蛋白生成：按长度/CATH/motif 条件采样 atom37 与残基类型。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/laproteina",intro:{title:"30 秒看懂模型",subtitle:"Proteina + configure_inference/predict_step，做条件蛋白生成或训练",suited:["flow 骨架生成","motif 条件采样","CATH 长度生成","训练 step 联调"],notFor:["序列确定性折叠","直接替代 AF 结构预测","缺 inf_cfg 就采样"],prepare:["cfg_exp","checkpoint","条件字段","nsamples/nres"],obtain:["atom37 坐标","residue type","训练 loss","样本列表"],steps:["匹配 nn.name","配置 autoencoder","加载 Proteina","configure_inference","predict_step"]},install:{banner:"已为你创建 LaProteina 环境配置任务",prompt:["请为 LaProteina 创建一个可以直接跑通的环境配置任务。","模型定位：PyTorch Lightning product-space flow matching，入口 Proteina.training_step / predict_step。","适合任务：Flow matching 骨架生成、Motif 条件蛋白采样、CATH 条件长度生成、Proteina 训练 step 联调。","典型输入：cfg_exp / ckpt / 条件 cath_code·x_motif / nsamples·nres","典型输出：atom37 / residue type / loss / 样本 list","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience LaProteina 入口（examples/biosciences/laproteina）与 onescience.models.laproteina.Proteina 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：Flow matching 骨架生成、Motif 条件蛋白采样、CATH 条件长度生成、Proteina 训练 step 联调。"].join(`
-`)},scenarios:[{id:"laproteina-gen",title:"Flow matching 骨架生成",summary:"按长度条件用 predict_step 采样蛋白骨架与残基类型",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：指定长度 de novo 骨架"},{id:"inputData",label:"ckpt / nres",placeholder:"例如：proteina.ckpt、nres、nsamples、inf_cfg",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 atom37 样本供设计性评估",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"LaProteina 无/长度条件生成",keyConfig:"configure_inference、nsamples、nres、product flow modalities",resources:"GPU；先小 nsamples 冒烟",deliverables:"atom37、residue type、长度一致性检查"},requiredData:["匹配 cfg_exp 与 checkpoint","inf_cfg","nsamples/nres"],expectedOutputs:"生成样本坐标/类型与质检",promptTemplate:["请基于 {{system}} 完成 LaProteina「Flow matching 骨架生成」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请构造 Proteina(cfg_exp)、加载 ckpt，调用 configure_inference(inf_cfg) 后 predict_step，","验证每个样本 atom37[n,37,3] 与 residue type[n] 长度一致，并规划设计性评估。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"laproteina-motif",title:"Motif 条件蛋白采样",summary:"在 x_motif 条件下做 scaffolding 式采样",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：功能 motif 支架"},{id:"inputData",label:"motif 坐标 / ckpt",placeholder:"例如：x_motif、mask、proteina ckpt、inf_cfg",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成保留 motif 的支架",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"LaProteina motif 条件采样",keyConfig:"x_motif、CFG unconditional 分支、autoguidance nn_ag",resources:"GPU；检查条件字段齐全",deliverables:"条件样本、motif 保留检查"},requiredData:["motif 坐标/掩码","支持条件的 cfg 与 ckpt","inf_cfg"],expectedOutputs:"motif 条件样本与保留度报告",promptTemplate:["请基于 {{system}} 做 LaProteina Motif 条件蛋白采样。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请保证 batch 含 x_motif（CFG 需要），configure_inference 后采样，","检查 motif 区域几何；条件失败时查字段而非删除断言。"].join(`
-`)},{id:"laproteina-cath",title:"CATH 条件长度生成",summary:"按 cath_code 与长度条件生成结构类别相关样本",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：某 CATH 超家族长度分布"},{id:"inputData",label:"cath_code / nres",placeholder:"例如：cath_code 列表、nres、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成类别条件骨架库",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"LaProteina CATH 条件生成",keyConfig:"cath_code、nres、nsamples、inf_cfg",resources:"GPU；按类别分批",deliverables:"分类别样本、长度统计"},requiredData:["cath_code 条件","长度设置","checkpoint"],expectedOutputs:"CATH 条件样本集与统计",promptTemplate:["请基于 {{system}} 完成 LaProteina CATH 条件长度生成。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请在 batch 中提供 cath_code 与 nres，采样后按类别汇总，","并说明如何序列化为结构文件供下游评估。"].join(`
-`)},{id:"laproteina-train",title:"Proteina 训练 step 联调",summary:"用 training_step 验证加噪、条件与损失可跑通",fields:t,capabilitySummary:{current:"LaProteina training_step 联调",keyConfig:"product_flowmatcher、nn.name、autoencoder、optimizer",resources:"GPU；先 1-2 step 冒烟",deliverables:"loss、日志、ckpt 路径"},requiredData:["完整 cfg_exp","训练 batch 字段","可选 autoencoder checkpoint"],expectedOutputs:"冒烟 loss 与训练入口",promptTemplate:["请基于 {{system}} 联调 LaProteina Proteina 训练 step。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 cfg_exp.nn.name 为源码支持值，配置 autoencoder 或显式 None，","跑 training_step；latent 维不匹配时核对 AE 与 flow matcher 配置。"].join(`
-`)}]},f={id:"medgemma",name:"MedGemma",category:"onemodel",domain:"bio",tags:["生信","医学LLM","问答","报告","多模态"],blurb:"医学文本/影像理解：Chat 风格消息经 vLLM 或 Transformers 做问答与报告辅助。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/medgemma",intro:{title:"30 秒看懂模型",subtitle:"MedGemmaPredictor 统一格式化 prompt，支持医学 QA、报告总结与批量评测",suited:["医学问答","报告辅助","MedQA 评测","医学提示生成"],notFor:["自动临床决策","蛋白/对接任务","未落地的图像 runner 当生产"],prepare:["model_path","tokenizer","JSON/JSONL messages","采样参数"],obtain:["OpenAI 兼容响应","预测文件","评测汇总","错误日志"],steps:["选 4B/27B","检查 vLLM","准备 JSONL","推理","人工复核"]},install:{banner:"已为你创建 MedGemma 环境配置任务",prompt:["请为 MedGemma 创建一个可以直接跑通的环境配置任务。","模型定位：医学多模态/文本模型，入口 MedGemmaPredictor + vLLMModelRunner/TransformersModelRunner。","适合任务：医学文本问答推理、报告辅助总结、MedQA 批量评测、vLLM 多模态医学提示。","典型输入：model_path / tokenizer / messages JSONL / temperature·max tokens","典型输出：Chat 响应 / 预测文件 / 评测表 / 错误目录","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience MedGemma 入口（examples/biosciences/medgemma）与 MedGemmaPredictor / vLLMModelRunner 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：医学文本问答推理、报告辅助总结、MedQA 批量评测、vLLM 多模态医学提示。"].join(`
-`)},scenarios:[{id:"medgemma-qa",title:"医学文本问答推理",summary:"对医学问题 messages 生成辅助回答（非自动诊断）",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：医学知识问答集"},{id:"inputData",label:"问题 JSONL / 模型路径",placeholder:"例如：qa.jsonl、medgemma-27b、temperature",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出需人工复核的回答",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MedGemma 医学问答",keyConfig:"model_path、temperature 偏低、max tokens、免责声明",resources:"GPU；优先 vLLM，失败回退 Transformers",deliverables:"回答文件、错误日志、复核标记"},requiredData:["含 messages/text/question 的 JSONL","本地模型与 tokenizer 路径","安全复核要求"],expectedOutputs:"问答输出与人工复核清单",promptTemplate:["请基于 {{system}} 完成 MedGemma「医学文本问答推理」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请加载本地权重（不可只写模型名），用 MedGemmaPredictor 推理；","医疗场景用较低 temperature，输出必须带人工复核说明，禁止作为自动诊断/处方。","vLLM 失败时设 use_vllm=false。隐私数据需去标识化。"].join(`
-`)},{id:"medgemma-report",title:"报告辅助总结",summary:"对医学报告文本做摘要/结构化辅助说明",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：放射/病理报告草稿"},{id:"inputData",label:"报告原文 / 提示",placeholder:"例如：reports.jsonl、总结提示词、模型路径",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成需医生确认的摘要",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MedGemma 报告辅助",keyConfig:"提示词约束、保留原文引用、版本与参数记录",resources:"GPU；批量写汇总文件",deliverables:"摘要、原文对照、参数记录"},requiredData:["报告文本 JSONL","模型路径","输出约束模板"],expectedOutputs:"辅助摘要与原文/参数归档",promptTemplate:["请基于 {{system}} 做 MedGemma 报告辅助总结。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请保留原文、提示词、模型版本与采样参数，生成结构化摘要供医生确认，","不得输出可执行处方或急救决策；空响应写入错误目录并支持重试。"].join(`
-`)},{id:"medgemma-medqa",title:"MedQA 批量评测",summary:"在 MedQA 等基准上批量推理并汇总准确率",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：MedQA 验证集"},{id:"inputData",label:"评测 JSONL / 模型",placeholder:"例如：medqa.jsonl、seed、temperature、model_path",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：可复现的准确率报告",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MedGemma 基准评测",keyConfig:"固定 seed/temperature/版本、汇总脚本",resources:"多卡可加速；结果需可复现",deliverables:"预测汇总、准确率、错误样本"},requiredData:["基准 JSONL","固定采样参数","本地权重"],expectedOutputs:"评测表与可复现配置卡",promptTemplate:["请基于 {{system}} 运行 MedGemma MedQA 批量评测。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请固定 seed、temperature 与模型版本，批量推理并统计准确率，","导出错误样本供分析；显存不足时降 max length 或换更小变体。"].join(`
-`)},{id:"medgemma-vl",title:"vLLM 多模态医学提示",summary:"尝试影像相关提示生成并标注多模态实现风险",fields:t,capabilitySummary:{current:"MedGemma 多模态提示（谨慎）",keyConfig:"4B 变体、use_vllm、图像 runner 可用性标注",resources:"GPU；源码图像能力不完整须标注风险",deliverables:"提示输出、风险说明、回退方案"},requiredData:["任务是否真正需要图像","4B 权重路径","可接受的回退（纯文本）"],expectedOutputs:"多模态试运行报告与回退路径",promptTemplate:["请基于 {{system}} 做 MedGemma vLLM 多模态医学提示试运行。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请选择 4B 变体并检查图像 runner 是否可用；若未完整实现，明确风险并回退文本描述输入。","任何输出仅供研究辅助，必须人工医学审阅。"].join(`
-`)}]},y={id:"molsculptor",name:"MolSculptor",category:"onemodel",domain:"bio",tags:["生信","小分子生成","药物设计","SMILES","latent-diffusion"],blurb:"JAX 小分子生成/优化：latent diffusion + reward/NSGA-II，产出 SMILES 与性质分数。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/molsculptor",intro:{title:"30 秒看懂模型",subtitle:"围绕 SMILES/分子图与 docking reward，做 de novo 生成、多目标优化或训练",suited:["SMILES 生成","多目标优化","docking reward 筛选","latent 训练"],notFor:["蛋白折叠/设计主任务","基因组建模","跨 case 混用 docking 路径"],prepare:["SMILES/case 配置","checkpoint","reward/receptor","采样策略"],obtain:["生成 SMILES","QED/SA/LogP","docking reward","Pareto 集"],steps:["确认小分子任务","选 de novo/优化","配 reward","采样","有效性过滤"]},install:{banner:"已为你创建 MolSculptor 环境配置任务",prompt:["请为 MolSculptor 创建一个可以直接跑通的环境配置任务。","模型定位：JAX/Flax 小分子 latent diffusion 工具链，入口 inference/training/case 脚本。","适合任务：SMILES 从头分子生成、多目标 NSGA-II 优化、Docking reward 筛选、Latent diffusion 训练。","典型输入：SMILES·分子图 / case 配置 / ckpt / LogP·QED·docking reward","典型输出：SMILES / 性质分数 / docking reward / 候选集","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience MolSculptor 入口（examples/biosciences/molsculptor）与 molsculptor inference/training/case 脚本 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：SMILES 从头分子生成、多目标 NSGA-II 优化、Docking reward 筛选、Latent diffusion 训练。"].join(`
-`)},scenarios:[{id:"molsculptor-denovo",title:"SMILES 从头分子生成",summary:"用 latent diffusion 从噪声或先验采样新 SMILES",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：先导化合物空间探索"},{id:"inputData",label:"ckpt / 采样配置",placeholder:"例如：diffusion ckpt、vocab、采样温度 top-k",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出有效 SMILES 库",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MolSculptor de novo 生成",keyConfig:"checkpoint、vocab、温度、top-k/top-p、padding",resources:"GPU；先小批量查有效性",deliverables:"SMILES、有效性率、去重列表"},requiredData:["匹配 ckpt/vocab/config","采样参数","输出目录"],expectedOutputs:"生成 SMILES 与有效性报告",promptTemplate:["请基于 {{system}} 完成 MolSculptor「SMILES 从头分子生成」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认任务为小分子而非蛋白/DNA，加载一致的 ckpt/vocab/config，","运行 inference 采样并检查 RDKit 有效性与重复率；无效过多时收紧温度或过滤。","如需安装 RDKit/OpenBabel，请先向我请求批准。"].join(`
-`)},{id:"molsculptor-nsga",title:"多目标 NSGA-II 优化",summary:"在性质多目标下用 NSGA-II 筛选 Pareto 分子",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：QED/SA/LogP 权衡"},{id:"inputData",label:"初始分子 / reward",placeholder:"例如：init SMILES pickle、reward 权重、NSGA 代数",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：交付 Pareto front 供人工选",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MolSculptor NSGA-II 优化",keyConfig:"多目标 reward、NSGA-II、相似度约束",resources:"GPU/CPU 混合；保留 front",deliverables:"Pareto 分子、目标分数、重复率"},requiredData:["初始分子集","reward 定义","优化代数/种群"],expectedOutputs:"Pareto front 与分数表",promptTemplate:["请基于 {{system}} 做 MolSculptor 多目标 NSGA-II 优化。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请配置 LogP/QED/SA/Tanimoto 等 reward 与 NSGA-II 参数，","输出 Pareto front 并去重；供人工或下游 ADMET 筛选。"].join(`
-`)},{id:"molsculptor-dock",title:"Docking reward 筛选",summary:"以 docking/case-local receptor 为 reward 筛选生成分子",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：针对靶点口袋的分子优化"},{id:"inputData",label:"receptor PDBQT / case",placeholder:"例如：case 目录、receptor、DSDP/docking 脚本、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：高 docking reward 候选",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"MolSculptor docking reward",keyConfig:"case-local 路径、receptor、docking 脚本、不可跨 case 混用",resources:"需 docking 依赖；不可用时退回性质 reward",deliverables:"docking reward 表、候选 SMILES"},requiredData:["case 配置与 receptor","docking 工具可用","生成/优化 ckpt"],expectedOutputs:"按 docking reward 排序的候选",promptTemplate:["请基于 {{system}} 完成 MolSculptor Docking reward 筛选。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请使用 case-local receptor/路径配置 docking reward，运行生成/优化，","检查 reward 分布；docking 不可用时先用 RDKit 性质 reward 并记录限制。"].join(`
-`)},{id:"molsculptor-train",title:"Latent diffusion 训练",summary:"训练或预训练 graph encoder/decoder 与 latent diffusion",fields:t,capabilitySummary:{current:"MolSculptor 训练",keyConfig:"autoencoder pretrain 或 diffusion training、图大小 padding",resources:"多卡 GPU；图过大需调 padding",deliverables:"checkpoint、训练曲线、验证 SMILES"},requiredData:["分子数据集","训练 config","依赖 RDKit 等"],expectedOutputs:"训练 ckpt 与验证采样",promptTemplate:["请基于 {{system}} 联调 MolSculptor Latent diffusion 训练。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请区分 autoencoder pretrain 与 diffusion training，检查分子图不超过 padding，","跑短训冒烟再正式训练；无 ckpt 时也可只产出数据与配置计划。"].join(`
-`)}]},g={id:"openfold",name:"OpenFold",category:"onemodel",domain:"bio",tags:["生信","结构预测","PyTorch","AF2","Evoformer"],blurb:"PyTorch AF2 风格：OpenFold feature dict 经 AlphaFold.forward 训练或结构推理。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/openfold",intro:{title:"30 秒看懂模型",subtitle:"用 openfold_data_pipeline 特征 + AlphaFold(config) 做推理/训练与长链 chunking",suited:["AF2 风格推理","ExtraMSA 训练","Recycling 校验","长链 chunking"],notFor:["FASTA 直喂 forward","AF3 多实体配体主路径","混用 AF/Protenix 特征"],prepare:["OpenFold config","feature dict","state dict","recycling 维"],obtain:["atom37","表征","辅助 heads","训练 loss"],steps:["确认模式","构造特征","加载 AlphaFold","forward","序列化/loss"]},install:{banner:"已为你创建 OpenFold 环境配置任务",prompt:["请为 OpenFold 创建一个可以直接跑通的环境配置任务。","模型定位：PyTorch AlphaFold2 风格结构模型，入口 onescience.models.openfold.model.AlphaFold。","适合任务：PyTorch AF2 结构推理、ExtraMSA/Template 训练前向、Recycling 特征校验、Chunking 长链推理。","典型输入：OpenFold config / feature dict（末维 recycling） / state dict","典型输出：atom37 / MSA·pair·single / loss / 辅助头","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience OpenFold 入口（examples/biosciences/openfold）与 onescience.models.openfold.model.AlphaFold 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：PyTorch AF2 结构推理、ExtraMSA/Template 训练前向、Recycling 特征校验、Chunking 长链推理。"].join(`
-`)},scenarios:[{id:"openfold-infer",title:"PyTorch AF2 结构推理",summary:"对 OpenFold feature dict 做 eval/no-grad 结构预测",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：需 PyTorch 栈的单体预测"},{id:"inputData",label:"特征 / config / 权重",placeholder:"例如：features.pt、openfold config、state dict",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 atom37 与置信头",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"OpenFold 推理",keyConfig:"globals/model config、recycling 末维、precision",resources:"GPU；先单样本",deliverables:"atom37、mask、辅助头字段"},requiredData:["OpenFold feature dict","匹配 config 与 state dict","device/precision"],expectedOutputs:"结构张量/PDB 与字段清单",promptTemplate:["请基于 {{system}} 完成 OpenFold「PyTorch AF2 结构推理」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用 openfold_data_pipeline 构造特征（勿把 FASTA 直传 forward），","实例化 AlphaFold(config) 加载 state dict，eval/no-grad 前向并只读实际存在的辅助头。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"openfold-train",title:"ExtraMSA/Template 训练前向",summary:"开启 Extra MSA/template 做训练前向与 loss",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：微调 OpenFold"},{id:"inputData",label:"训练 batch / 损失",placeholder:"例如：train features、labels、template/ExtraMSA 开关",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证 autograd 与 loss",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"OpenFold 训练前向",keyConfig:"template、Extra MSA、autograd、activation checkpointing",resources:"多卡 GPU；注意显存",deliverables:"loss、梯度冒烟、日志"},requiredData:["训练 feature dict 与标签","训练 config","损失函数"],expectedOutputs:"loss 与训练入口说明",promptTemplate:["请基于 {{system}} 联调 OpenFold ExtraMSA/Template 训练前向。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 config 中 template/Extra MSA 与特征一致，保持 autograd，","model(batch) 后接项目损失；缺字段回 data pipeline，勿伪造。"].join(`
-`)},{id:"openfold-recycle",title:"Recycling 特征校验",summary:"检查所有 tensor 末维 recycling 与轮数配置一致",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：特征管线质检"},{id:"inputData",label:"feature dict",placeholder:"例如：features、预期 recycle 数、monomer/multimer 模式",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：修复 shape 后再推理",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"OpenFold recycling 校验",keyConfig:"batch 末维=recycling、模式与 checkpoint 一致",resources:"本地 CPU 即可检查",deliverables:"校验报告、修复建议"},requiredData:["feature dict","config.globals 模式","预期 recycle"],expectedOutputs:"recycling 一致性报告",promptTemplate:["请为 {{system}} 做 OpenFold Recycling 特征校验。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请逐项检查特征末维与 monomer/multimer/seqemb 模式，","输出不匹配字段清单并回到 openfold_data_pipeline 修复。"].join(`
-`)},{id:"openfold-chunk",title:"Chunking 长链推理",summary:"对长序列启用 chunking/checkpointing 控制显存",fields:t,capabilitySummary:{current:"OpenFold 长链 chunking",keyConfig:"chunking、checkpointing、减小 MSA/template",resources:"大显存 GPU；先单样本预检",deliverables:"结构结果、显存策略记录"},requiredData:["长序列 feature dict","显存预算","chunk/ckpt 配置"],expectedOutputs:"长链推理结果与资源策略",promptTemplate:["请基于 {{system}} 完成 OpenFold Chunking 长链推理。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请启用源码 chunking/checkpointing，必要时减小 MSA/template，","先单样本显存预检再正式推理；OOM 时记录可复现的资源参数。"].join(`
-`)}]},b={id:"proteinmpnn",name:"ProteinMPNN",category:"onemodel",domain:"bio",tags:["生信","逆折叠","序列设计","骨架条件","蛋白设计"],blurb:"骨架条件序列设计：ProteinMPNN 对 N/CA/C/O 骨架采样或打分氨基酸序列。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/ProteinMPNN",intro:{title:"30 秒看懂模型",subtitle:"parse_PDB + tied_featurize 后 sample/tied_sample/forward，做逆折叠设计与约束",suited:["骨架序列设计","tied 约束采样","固定位点打分","CA-only 对照"],notFor:["从序列预测结构","无骨架输入","破坏 chain_M 固定约束"],prepare:["骨架 PDB","设计/固定 mask","checkpoint","temperature"],obtain:["设计序列","log probability","采样诊断","约束检查"],steps:["解析骨架","tied_featurize","加载 ProteinMPNN","sample/打分","结构验证交接"]},install:{banner:"已为你创建 ProteinMPNN 环境配置任务",prompt:["请为 ProteinMPNN 创建一个可以直接跑通的环境配置任务。","模型定位：蛋白逆折叠序列模型，入口 ProteinMPNN.sample / tied_sample / forward。","适合任务：骨架条件序列设计、Tied-position 约束采样、固定位点序列打分、CA-only 逆折叠对照。","典型输入：骨架 PDB / chain_M·fixed·omit-AA·PSSM / full 或 CA-only ckpt","典型输出：设计 FASTA / logprob / 采样诊断 / 约束核验","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience ProteinMPNN 入口（examples/biosciences/ProteinMPNN）与 protein_mpnn_utils.ProteinMPNN 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：骨架条件序列设计、Tied-position 约束采样、固定位点序列打分、CA-only 逆折叠对照。"].join(`
-`)},scenarios:[{id:"proteinmpnn-design",title:"骨架条件序列设计",summary:"对 RFdiffusion 等骨架采样氨基酸序列",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如： binder 骨架序列化"},{id:"inputData",label:"backbone PDB / ckpt",placeholder:"例如：backbone.pdb、设计链、temperature、num sequences",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 FASTA 供结构验证",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProteinMPNN sample 设计",keyConfig:"chain_M、temperature、full-backbone 原子序 N-CA-C-O",resources:"GPU/CPU；批量骨架可并行",deliverables:"FASTA、logprob、约束核验"},requiredData:["骨架 PDB","设计链与固定位点","匹配 checkpoint"],expectedOutputs:"设计序列 FASTA 与打分摘要",promptTemplate:["请基于 {{system}} 完成 ProteinMPNN「骨架条件序列设计」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请 parse_PDB → tied_featurize 构造 X/S/mask/chain_M 等，加载 ProteinMPNN，","调用 sample 生成序列并核验固定位点；输出交给 SimpleFold/OpenFold/AF3 验证。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"proteinmpnn-tied",title:"Tied-position 约束采样",summary:"对对称/绑定位点用 tied_sample 联合解码",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：对称多聚体设计"},{id:"inputData",label:"tied 位点 / PDB",placeholder:"例如：tied_positions.json、backbone.pdb、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：对称位点氨基酸一致",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProteinMPNN tied_sample",keyConfig:"tied positions、bias、decoding order",resources:"GPU；先小样本检查对称性",deliverables:"tied 序列、约束满足率"},requiredData:["骨架与 tied 定义","checkpoint","采样温度"],expectedOutputs:"tied 设计序列与满足率报告",promptTemplate:["请基于 {{system}} 做 ProteinMPNN Tied-position 约束采样。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请在 featurize 中加入 tied 张量并调用 tied_sample，","核验绑定位点氨基酸一致；违反约束时检查 chain_M 与 tied 特征。"].join(`
-`)},{id:"proteinmpnn-score",title:"固定位点序列打分",summary:"对已知序列在骨架上计算 log probability",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：突变序列适配性"},{id:"inputData",label:"序列 / 骨架",placeholder:"例如：seq.fasta、backbone.pdb、fixed positions",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：比较野生型与突变 logprob",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProteinMPNN forward 打分",keyConfig:"forward logprob、loss_nll/loss_smoothed、固定 mask",resources:"CPU/GPU 均可",deliverables:"logprob 表、位点贡献"},requiredData:["骨架坐标","待打分序列","checkpoint"],expectedOutputs:"序列 logprob 与比较表",promptTemplate:["请基于 {{system}} 完成 ProteinMPNN 固定位点序列打分。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请构造特征后调用 forward 得到 (B,L,21) logprob，","按需 loss_nll，输出序列/位点分数供排序。"].join(`
-`)},{id:"proteinmpnn-ca",title:"CA-only 逆折叠对照",summary:"在 CA-only 模式与配套 ckpt 下设计或打分",fields:t,capabilitySummary:{current:"ProteinMPNN CA-only",keyConfig:"CA-only 配置与 checkpoint 必须配套",resources:"GPU/CPU",deliverables:"CA-only 序列、与全原子对照"},requiredData:["CA 骨架","CA-only ckpt","设计参数"],expectedOutputs:"CA-only 设计结果与对照说明",promptTemplate:["请基于 {{system}} 做 ProteinMPNN CA-only 逆折叠对照。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 CA-only 配置与 checkpoint 配套，勿混用全原子权重；","输出序列并建议与 full-backbone 结果及结构验证对照。"].join(`
-`)}]},D={id:"protenix",name:"Protenix",category:"onemodel",domain:"bio",tags:["生信","复合物","AF3-style","扩散","Pairformer"],blurb:"PyTorch AF3 风格复合物模型：input_feature_dict 上 train/inference/eval 扩散预测。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/protenix",intro:{title:"30 秒看懂模型",subtitle:"Protenix.forward 按 mode 组合 Pairformer/扩散头，做复合物推理或训练评估",suited:["AF3-style 推理","扩散训练","配体-核酸验证","对称置换评估"],notFor:["JSON/FASTA 直喂","与 AF3 JAX 混用 batch","train 省略 labels"],prepare:["configs","input_feature_dict","checkpoint","mode 与 labels"],obtain:["pred_dict","confidence","contact prob","log_dict"],steps:["protenix_data_pipeline","构造 Protenix","按 mode forward","解析坐标","排序"]},install:{banner:"已为你创建 Protenix 环境配置任务",prompt:["请为 Protenix 创建一个可以直接跑通的环境配置任务。","模型定位：OneScience AF3-style 复合物模型，入口 Protenix(configs).forward(mode=...)。","适合任务：AF3-style 复合物推理、Pairformer 扩散训练、配体-核酸联合验证、SymmetricPermutation 评估。","典型输入：configs / input_feature_dict / labels·SymmetricPermutation / checkpoint","典型输出：pred_dict / confidence / contact / log_dict","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience Protenix 入口（examples/biosciences/protenix）与 onescience.models.protenix.protenix.Protenix 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：AF3-style 复合物推理、Pairformer 扩散训练、配体-核酸联合验证、SymmetricPermutation 评估。"].join(`
-`)},scenarios:[{id:"protenix-infer",title:"AF3-style 复合物推理",summary:"mode=inference 对 feature dict 预测复合物坐标与置信度",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：蛋白-配体复合物验证"},{id:"inputData",label:"feature dict / ckpt",placeholder:"例如：input_feature_dict、protenix.ckpt、configs",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出坐标与 confidence",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Protenix inference",keyConfig:"mode=inference、labels=None、diffusion/confidence 头",resources:"大显存 GPU；可调 seed/cycle",deliverables:"coordinates、confidence、contact probabilities"},requiredData:["protenix_data_pipeline 特征","匹配 configs/ckpt","device/precision"],expectedOutputs:"预测结构与置信度产物",promptTemplate:["请基于 {{system}} 完成 Protenix「AF3-style 复合物推理」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请召回 protenix_data_pipeline 构造 input_feature_dict（勿直喂 JSON/FASTA），","Protenix(configs) 加载权重后 mode='inference' forward，解析 pred_dict 坐标与置信度。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"protenix-train",title:"Pairformer 扩散训练",summary:"mode=train 带 labels 与 SymmetricPermutation 训练",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：复合物结构微调"},{id:"inputData",label:"特征 / 标签 / step",placeholder:"例如：feature、label_full/label_dict、current_step、permutation",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证 train loss 可反传",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Protenix train 模式",keyConfig:"mode=train、label_full_dict、SymmetricPermutation、current_step",resources:"多卡；注意断言与显存",deliverables:"loss/log_dict、更新标签、耗时"},requiredData:["训练特征与完整/裁剪标签","SymmetricPermutation","configs"],expectedOutputs:"训练日志与 checkpoint 入口",promptTemplate:["请基于 {{system}} 联调 Protenix Pairformer 扩散训练。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请设置 model.training，准备 labels 与 SymmetricPermutation，mode='train' 调用；","断言失败时检查 mode/labels/permutation，勿省略必填参数。"].join(`
-`)},{id:"protenix-ligand",title:"配体-核酸联合验证",summary:"对含配体/核酸的复合物做推理并检查接触",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：RNP+小分子体系"},{id:"inputData",label:"多组分特征",placeholder:"例如：protein+NA+ligand features、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证跨分子接触与置信度",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"Protenix 多组分验证",keyConfig:"token/atom 特征、ligand/NA 映射、contact prob",resources:"大显存；可降 diffusion batch",deliverables:"复合物结构、接触与失败诊断"},requiredData:["含配体/核酸的 feature dict","checkpoint","实体映射"],expectedOutputs:"多组分验证报告",promptTemplate:["请基于 {{system}} 做 Protenix 配体-核酸联合验证。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 atom/token/pair 特征完整，推理后检查配体/核酸几何与 contact probabilities，","低置信区域建议与 AlphaFold3 JAX 路线交叉验证（注意不共享 batch adapter）。"].join(`
-`)},{id:"protenix-eval",title:"SymmetricPermutation 评估",summary:"mode=eval 在对称置换下评估预测与标签",fields:t,capabilitySummary:{current:"Protenix eval 模式",keyConfig:"mode=eval、可选真值标签、置换对象",resources:"GPU；与训练同规格特征",deliverables:"eval 指标、log_dict、置换日志"},requiredData:["eval feature dict","可选 label_dict","SymmetricPermutation"],expectedOutputs:"评估指标与日志",promptTemplate:["请基于 {{system}} 完成 Protenix SymmetricPermutation 评估。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按是否有真值准备 label，mode='eval' forward，","记录置换与指标；显存不足时降低采样/循环并启用 chunking。"].join(`
-`)}]},k={id:"protoken",name:"ProToken",category:"onemodel",domain:"bio",tags:["生信","结构分词","VQ","重建","PT-DiT支撑"],blurb:"蛋白结构 VQ tokenizer：PDB↔结构 token，为 PT-DiT 提供词表与解码后端。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/flax_models/protoken",intro:{title:"30 秒看懂模型",subtitle:"infer.py 编码 PDB，decode_structure.py 从 token 重建，服务 PT-DiT 联动",suited:["结构 token 编码","token 重建 PDB","PT-DiT 词表对齐","重建质检"],notFor:["独立 AF 折叠","ProteinMPNN 逆折叠","序列直折"],prepare:["PDB 或 vq indexes","ProToken ckpt","embedding","padding_len"],obtain:["vq_code_indexes.pkl","重建 PDB","aux","版本记录"],steps:["选 encode/decode","检查 padding","跑脚本","几何 QC","交 PT-DiT"]},install:{banner:"已为你创建 ProToken 环境配置任务",prompt:["请为 ProToken 创建一个可以直接跑通的环境配置任务。","模型定位：结构离散 VQ tokenizer，入口 infer.py / decode_structure.py（flax_models/protoken）。","适合任务：PDB 结构 token 编码、VQ token 重建 PDB、PT-DiT 词表对齐、Codebook 重建质检。","典型输入：PDB 或 vq_code_indexes / ProToken ckpt·embedding / padding_len","典型输出：vq_code_indexes.pkl / 重建 PDB / aux / 版本卡","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience ProToken 入口（src/onescience/flax_models/protoken）与 infer.py / decode_structure.py 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：PDB 结构 token 编码、VQ token 重建 PDB、PT-DiT 词表对齐、Codebook 重建质检。"].join(`
-`)},scenarios:[{id:"protoken-encode",title:"PDB 结构 token 编码",summary:"将 PDB/atom features 编码为 VQ 结构 token",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：设计骨架入库"},{id:"inputData",label:"PDB / ckpt",placeholder:"例如：input.pdb、protoken.ckpt、padding_len",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 vq_code_indexes.pkl",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProToken encode",keyConfig:"padding_len≥残基数、codebook 版本、infer.py",resources:"GPU/CPU；先短链验证",deliverables:"vq_code_indexes.pkl、input features、aux"},requiredData:["PDB","ProToken checkpoint","足够 padding_len"],expectedOutputs:"token indexes 与编码日志",promptTemplate:["请基于 {{system}} 完成 ProToken「PDB 结构 token 编码」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 padding_len 覆盖残基数，运行 infer.py，","检查 token 长度与 aux；残基过长时提高 padding 或切分。记录 codebook 版本供 PT-DiT。"].join(`
-`)},{id:"protoken-decode",title:"VQ token 重建 PDB",summary:"从 vq_code_indexes 经 decoder 重建 atom37/PDB",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：PT-DiT 输出解码"},{id:"inputData",label:"token indexes / decoder",placeholder:"例如：vq_code_indexes.pkl、decoder ckpt、padding_len",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：重建 PDB 做几何检查",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProToken decode",keyConfig:"decode_structure.py、embedding、codebook 一致",resources:"GPU；解码后几何 QC",deliverables:"重建 PDB、QC 报告"},requiredData:["vq_code_indexes","匹配 decoder/embedding","padding_len"],expectedOutputs:"重建 PDB 与几何质检",promptTemplate:["请基于 {{system}} 做 ProToken VQ token 重建 PDB。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用 decode_structure.py 解码，检查链/缺失原子与 codebook 一致性，","异常时保留 indexes 并诊断 PDB 质量或版本不匹配。"].join(`
-`)},{id:"protoken-align",title:"PT-DiT 词表对齐",summary:"核对 ProToken 词表/embedding 与 PT-DiT checkpoint 配套",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：准备 co-design 资源"},{id:"inputData",label:"ProToken / PT-DiT 资源",placeholder:"例如：protoken_emb、aatype_emb、两端 ckpt、nres",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：确认可联合采样",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ProToken↔PT-DiT 对齐",keyConfig:"nres 与 padding 一致、embedding/ckpt 版本",resources:"本地检查 + 最小联动冒烟",deliverables:"对齐报告、冒烟结果"},requiredData:["ProToken ckpt/emb","PT-DiT ckpt","nres/padding 约定"],expectedOutputs:"资源对齐卡与联动入口",promptTemplate:["请为 {{system}} 做 ProToken 与 PT-DiT 词表对齐。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请核对 codebook、embedding、nres/padding 与两端 checkpoint 版本，","跑最小 encode→PT-DiT→decode 冒烟；不匹配则停止生成并列出修复项。"].join(`
-`)},{id:"protoken-qc",title:"Codebook 重建质检",summary:"抽样 encode→decode 评估重建误差与可视化",fields:t,capabilitySummary:{current:"ProToken 重建质检",keyConfig:"抽样策略、几何指标、版本记录",resources:"可并行 CPU/GPU",deliverables:"QC 表、失败结构列表"},requiredData:["PDB 集合","ProToken 资源","质检阈值"],expectedOutputs:"质检报告与失败案例",promptTemplate:["请基于 {{system}} 运行 ProToken Codebook 重建质检。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请小规模抽样 encode/decode，统计几何误差与失败模式，","形成是否可进入 PT-DiT 或结构验证的结论。"].join(`
-`)}]},S={id:"pt_dit",name:"PT-DiT",category:"onemodel",domain:"bio",tags:["生信","蛋白设计","扩散Transformer","co-design","ProToken"],blurb:"基于 ProToken 的序列-结构协同扩散：de novo/RePaint/插值生成 token 与氨基酸。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/pt_dit",intro:{title:"30 秒看懂模型",subtitle:"de_novo_design.py 采样 latent/protoken/aatype，并可经 ProToken 解码为 PDB",suited:["序列-结构协同生成","de novo 采样","RePaint 重绘","latent 插值"],notFor:["普通 FASTA 折叠","替代 ProteinMPNN","资源版本不一致"],prepare:["PT-DiT ckpt","ProToken 资源","nres/nsample","timesteps"],obtain:["result.pkl","protoken/aatype indexes","可选 PDB","候选 ID"],steps:["核对资源版本","设 nres","de novo/RePaint","解码","结构验证"]},install:{banner:"已为你创建 PT-DiT 环境配置任务",prompt:["请为 PT-DiT 创建一个可以直接跑通的环境配置任务。","模型定位：ProToken+AA embedding 的 Diffusion Transformer，入口 de_novo_design.py。","适合任务：序列-结构协同生成、De novo ProToken 采样、RePaint 局部重绘、Latent 插值演化。","典型输入：PT-DiT ckpt / ProToken ckpt·emb / nres·nsample / timesteps","典型输出：result.pkl / protoken·aatype indexes / 可选 PDB","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience PT-DiT 入口（examples/biosciences/pt_dit）与 de_novo_design.py + ProToken decoder 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：序列-结构协同生成、De novo ProToken 采样、RePaint 局部重绘、Latent 插值演化。"].join(`
-`)},scenarios:[{id:"pt_dit-codesign",title:"序列-结构协同生成",summary:"联合采样结构 token 与氨基酸 indexes 做 co-design",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：新蛋白候选库"},{id:"inputData",label:"两端 ckpt / nres",placeholder:"例如：pt_dit.ckpt、protoken 资源、nres、nsample_per_device",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出可解码的设计包",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"PT-DiT co-design",keyConfig:"protoken_emb+aatype_emb、diffusion timesteps、设备整除 batch",resources:"多卡；batch 需能被设备数整除",deliverables:"result.pkl、indexes、候选 ID"},requiredData:["版本一致的 PT-DiT 与 ProToken 资源","nres 满足 flash/padding","采样规模"],expectedOutputs:"协同生成产物与版本记录",promptTemplate:["请基于 {{system}} 完成 PT-DiT「序列-结构协同生成」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认 ProToken/AA embedding 与 PT-DiT ckpt 配套，设置 nres 与 timesteps，","运行 de_novo_design（或等价），保存 latent 与 indexes；勿把普通 FASTA/OpenFold batch 送入。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"pt_dit-denovo",title:"De novo ProToken 采样",summary:"从头采样 protoken/aatype 并可选解码 PDB",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：指定长度新折叠"},{id:"inputData",label:"采样参数",placeholder:"例如：nres、nsample、是否 decode_structures",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：得到 PDB 候选供折叠复核",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"PT-DiT de novo",keyConfig:"de_novo_design.py、decode 开关、nsample_per_device",resources:"GPU；显存紧时先不解码",deliverables:"indexes、可选 PDB、质检"},requiredData:["配套资源","nres/nsample","是否立即 ProToken decode"],expectedOutputs:"de novo 设计包与可选 PDB",promptTemplate:["请基于 {{system}} 做 PT-DiT De novo ProToken 采样。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请运行 de novo 采样；解码失败时先保存 token/latent。","候选用 SimpleFold/OpenFold 单体验证或 AF3/Protenix 复合物验证。"].join(`
-`)},{id:"pt_dit-repaint",title:"RePaint 局部重绘",summary:"对局部结构/序列 token 做 RePaint 式重采样",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：活性位点邻域重设计"},{id:"inputData",label:"条件 mask / 初始 indexes",placeholder:"例如：已知 protoken indexes、重绘区域 mask、timesteps",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：局部更新且保留框架",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"PT-DiT RePaint",keyConfig:"repaint mask、条件 indexes、timesteps",resources:"GPU；保留未掩码区域检查",deliverables:"重绘 indexes、差异报告、可选 PDB"},requiredData:["初始 token/序列","重绘区域定义","配套 ckpt"],expectedOutputs:"RePaint 结果与保留区核验",promptTemplate:["请基于 {{system}} 完成 PT-DiT RePaint 局部重绘。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请定义 repaint 区域与条件，采样后核对未掩码区保持，","输出更新 token/序列并规划结构验证。"].join(`
-`)},{id:"pt_dit-interp",title:"Latent 插值演化",summary:"在 latent 空间插值生成过渡序列-结构",fields:t,capabilitySummary:{current:"PT-DiT latent interpolation",keyConfig:"端点 latent、插值步数、解码策略",resources:"GPU；轨迹可只存 indexes",deliverables:"插值轨迹、可选 PDB 序列"},requiredData:["两端 latent 或 indexes","插值参数","资源版本"],expectedOutputs:"插值轨迹与抽样结构",promptTemplate:["请基于 {{system}} 做 PT-DiT Latent 插值演化。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请在配套资源下对端点 latent 插值，保存轨迹 indexes，","抽样解码并做几何/折叠检查，保留三类产物便于回溯。"].join(`
-`)}]},P={id:"rfdiffusion",name:"RFdiffusion",category:"onemodel",domain:"bio",tags:["生信","骨架生成","扩散","motif","binder"],blurb:"蛋白骨架扩散设计：contig/热点/部分扩散生成 backbone，供 ProteinMPNN 接续。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/RFdiffusion",intro:{title:"30 秒看懂模型",subtitle:"run_inference.py + Hydra contig，覆盖无条件、motif、binder 与 partial diffusion",suited:["无条件骨架","motif scaffolding","binder 设计","partial diffusion"],notFor:["FASTA 折叠主任务","已有序列直接验证","编造 binder 热点"],prepare:["目标/motif PDB","contig","checkpoint","num_designs"],obtain:["骨架 PDB","TRB","轨迹","设计 ID"],steps:["选模式","写 contig","跑推理","检查 TRB","交 ProteinMPNN"]},install:{banner:"已为你创建 RFdiffusion 环境配置任务",prompt:["请为 RFdiffusion 创建一个可以直接跑通的环境配置任务。","模型定位：RoseTTAFold 风格骨架扩散，入口 run_inference.py（Hydra contig/hotspot）。","适合任务：无条件骨架扩散生成、Motif scaffolding 设计、Binder 界面骨架设计、Partial diffusion 改造。","典型输入：PDB·contig·hotspot / checkpoint / num_designs / 输出前缀","典型输出：骨架 PDB / TRB / 轨迹 / 日志","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience RFdiffusion 入口（examples/biosciences/RFdiffusion）与 run_inference.py 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：无条件骨架扩散生成、Motif scaffolding 设计、Binder 界面骨架设计、Partial diffusion 改造。"].join(`
-`)},scenarios:[{id:"rfdiffusion-uncond",title:"无条件骨架扩散生成",summary:"按长度 contig 无条件采样蛋白骨架",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：新折叠长度探索"},{id:"inputData",label:"长度 contig / ckpt",placeholder:"例如：contig='100-150'、num_designs、output_prefix",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出 backbone 供序列设计",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"RFdiffusion 无条件生成",keyConfig:"contig 长度、num_designs、checkpoint、是否存轨迹",resources:"GPU；先少 designs 冒烟",deliverables:"骨架 PDB、TRB、日志"},requiredData:["长度 contig","checkpoint 目录","输出前缀"],expectedOutputs:"骨架文件与下游 ProteinMPNN 入口",promptTemplate:["请基于 {{system}} 完成 RFdiffusion「无条件骨架扩散生成」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用 run_inference.py 写出 Hydra contig/长度与 num_designs，","检查 PDB/TRB；成功后规划 ProteinMPNN 序列设计与结构验证。","如需安装依赖或提交作业，请先向我请求批准。"].join(`
-`)},{id:"rfdiffusion-motif",title:"Motif scaffolding 设计",summary:"围绕输入 motif 生成支架骨架并检查 RMSD",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：功能 motif 支架"},{id:"inputData",label:"motif PDB / contig",placeholder:"例如：motif.pdb、contig 固定片段语法、链残基编号",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：motif 保留且支架多样",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"RFdiffusion motif scaffolding",keyConfig:"contig 固定/可变段、链 ID 与残基编号一致",resources:"GPU；核对 PDB 编号",deliverables:"scaffold PDB、motif RMSD、TRB"},requiredData:["motif PDB","正确 contig 语法","checkpoint"],expectedOutputs:"scaffold 与 motif 保留报告",promptTemplate:["请基于 {{system}} 做 RFdiffusion Motif scaffolding 设计。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请把 motif 约束写成 contig（链 ID/残基/断链语法一致），采样后检查 motif RMSD；","解析失败时重核编号与异质原子清理。"].join(`
-`)},{id:"rfdiffusion-binder",title:"Binder 界面骨架设计",summary:"针对靶标链与热点生成 binder 骨架",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：靶蛋白界面 binder"},{id:"inputData",label:"target PDB / hotspot",placeholder:"例如：target.pdb、hotspot 残基、binder 长度范围",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：界面候选骨架库",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"RFdiffusion binder design",keyConfig:"靶标链、hotspot（不可编造）、binder 长度、num_designs",resources:"GPU；界面任务后续需复合物验证",deliverables:"binder 骨架、TRB、热点记录"},requiredData:["target PDB","明确热点或声明未指定","binder contig"],expectedOutputs:"binder 骨架与下游验证计划",promptTemplate:["请基于 {{system}} 完成 RFdiffusion Binder 界面骨架设计。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请明确靶标链、binder 长度与热点（不可凭空编造），运行采样，","输出骨架后接 ProteinMPNN，再用 AF3/Protenix 做复合物与界面筛选。"].join(`
-`)},{id:"rfdiffusion-partial",title:"Partial diffusion 改造",summary:"保留部分结构并对指定区域做部分扩散",fields:t,capabilitySummary:{current:"RFdiffusion partial diffusion",keyConfig:"输入长度与设计长度映射、保留区、扩散步",resources:"GPU；长度映射必须明确",deliverables:"改造骨架、保留区 RMSD、TRB"},requiredData:["输入 PDB","保留/扩散区定义","长度映射"],expectedOutputs:"partial 设计结果与保留核验",promptTemplate:["请基于 {{system}} 做 RFdiffusion Partial diffusion 改造。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请确认输入与设计长度映射，设置 partial diffusion 参数，","采样后检查保留区；资源不足时减少 num_designs 或关轨迹。"].join(`
-`)}]},C={id:"simplefold",name:"SimpleFold",category:"onemodel",domain:"bio",tags:["生信","结构预测","flow-matching","快速折叠","FoldingDiT"],blurb:"Flow matching 生成式折叠：处理后特征经 FoldingDiT/SimpleFold 采样蛋白结构。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/simplefold",intro:{title:"30 秒看懂模型",subtitle:"simplefold_data_pipeline + SimpleFold.predict_step，适合快速单体结构采样",suited:["flow 快速折叠","FoldingDiT 采样","ESM 特征准备","EMA 训练"],notFor:["FASTA 直喂模型类","复杂多实体配体主路径","ESM 维与 ckpt 不一致"],prepare:["处理后 feats","ckpt/EMA","sampler 配置","esm_model 匹配"],obtain:["采样坐标","可选 pLDDT","loss","EMA 状态"],steps:["跑 data pipeline","匹配 esm_s","加载 SimpleFold","predict_step","质量评估"]},install:{banner:"已为你创建 SimpleFold 环境配置任务",prompt:["请为 SimpleFold 创建一个可以直接跑通的环境配置任务。","模型定位：PyTorch Lightning flow-matching 折叠，入口 SimpleFold.predict_step / FoldingDiT。","适合任务：Flow matching 快速折叠、FoldingDiT 采样推理、ESM 特征管线准备、EMA checkpoint 训练。","典型输入：处理后 feats·esm_s / architecture·sampler / EMA ckpt","典型输出：采样结构 / 可选 pLDDT / loss / EMA","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience SimpleFold 入口（examples/biosciences/simplefold）与 onescience.models.simplefold.SimpleFold / FoldingDiT 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：Flow matching 快速折叠、FoldingDiT 采样推理、ESM 特征管线准备、EMA checkpoint 训练。"].join(`
-`)},scenarios:[{id:"simplefold-fold",title:"Flow matching 快速折叠",summary:"用 predict_step 从特征采样单体结构作初筛",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：设计序列快速过筛"},{id:"inputData",label:"特征 / EMA ckpt",placeholder:"例如：processed feats、simplefold.ckpt、sampler",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：快速 PDB 供高阶模型复核",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"SimpleFold 采样折叠",keyConfig:"processor、path、sampler、EMA checkpoint",resources:"GPU；适合单体快筛",deliverables:"采样坐标、可选置信度、日志"},requiredData:["simplefold_data_pipeline 特征","匹配 ckpt 组件","sampler 配置"],expectedOutputs:"结构文件与初筛结论",promptTemplate:["请基于 {{system}} 完成 SimpleFold「Flow matching 快速折叠」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请经 simplefold_data_pipeline 准备特征（FASTA 须在模型外转换），","加载 SimpleFold 与 EMA ckpt，predict_step 采样；高价值候选再用 OpenFold/Protenix 复核。"].join(`
-`)},{id:"simplefold-dit",title:"FoldingDiT 采样推理",summary:"直接调试 FoldingDiT.forward(noised_pos,t,feats) 采样表示",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：主干冒烟或研究采样"},{id:"inputData",label:"noised_pos / feats",placeholder:"例如：noised_pos、t、feats、self_cond、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：验证 predict_velocity 与 latent",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"FoldingDiT 主干推理",keyConfig:"noised_pos、t、feats、self_cond",resources:"GPU；先单步前向",deliverables:"predict_velocity、token latent"},requiredData:["与 esm_model 匹配的 feats","主干权重","噪声与时间步"],expectedOutputs:"主干输出与 shape 报告",promptTemplate:["请基于 {{system}} 做 SimpleFold FoldingDiT 采样推理调试。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请调用 FoldingDiT(noised_pos, t, feats, self_cond)，","检查 atom/token 映射与 ESM 通道；shape 失败回 data pipeline。"].join(`
-`)},{id:"simplefold-esmfeat",title:"ESM 特征管线准备",summary:"准备与所选 esm_model 匹配的 esm_s 及原子/token 特征",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：新序列进入 SimpleFold"},{id:"inputData",label:"FASTA / esm_model",placeholder:"例如：seqs.fasta、esm 权重、processor 配置",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：产出可 predict 的 feats",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"SimpleFold ESM 特征准备",keyConfig:"esm_model 与 esm_s 通道一致、atom-to-token mapping",resources:"GPU 提特征；可缓存",deliverables:"feats 包、映射检查、数据卡"},requiredData:["蛋白 FASTA","指定 esm_model","processor 配置"],expectedOutputs:"处理后特征与一致性卡",promptTemplate:["请为 {{system}} 准备 SimpleFold ESM 特征管线。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 checkpoint 的 esm_model 提取 esm_s，完成 atom/token 特征与 mask，","维度失败时恢复 ckpt 对应层数，不混用其它 ESM 变体。"].join(`
-`)},{id:"simplefold-ema",title:"EMA checkpoint 训练",summary:"用 training_step 与 EMA 训练/微调 SimpleFold",fields:t,capabilitySummary:{current:"SimpleFold 训练+EMA",keyConfig:"training_step、EMA、loss、scheduler",resources:"多卡 Lightning；先短训",deliverables:"EMA ckpt、loss 曲线"},requiredData:["训练特征协议","architecture/processor/loss/path/sampler","优化器配置"],expectedOutputs:"EMA checkpoint 与训练日志",promptTemplate:["请基于 {{system}} 联调 SimpleFold EMA checkpoint 训练。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请组装 SimpleFold 组件跑 training_step 与 EMA，","短训冒烟后正式训练；改 feature schema 时同步更新 processor 与测试。"].join(`
-`)}]},M={id:"targetdiff",name:"TargetDiff",category:"onemodel",domain:"bio",tags:["生信","口袋生成","配体扩散","ScorePosNet3D","分子生成"],blurb:"口袋条件 3D 配体扩散：ScorePosNet3D 联合采样坐标与原子类型，可接性质头。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/biosciences/targetdiff",intro:{title:"30 秒看懂模型",subtitle:"protein/ligand 张量上 train、sample_diffusion 或 likelihood，生成口袋配体",suited:["口袋配体生成","反向扩散采样","似然估计","性质预测头"],notFor:["仅对接已知配体","把坐标当完整化学分子","特征维与 ckpt 不一致"],prepare:["口袋原子特征","ligand 初始化","config/ckpt","num_steps"],obtain:["配体坐标/类型","轨迹","loss/likelihood","性质预测"],steps:["dataset 特征化","构造 ScorePosNet3D","train 或 sample","成键后处理","打分"]},install:{banner:"已为你创建 TargetDiff 环境配置任务",prompt:["请为 TargetDiff 创建一个可以直接跑通的环境配置任务。","模型定位：口袋条件配体扩散，入口 ScorePosNet3D.sample_diffusion / get_diffusion_loss。","适合任务：口袋条件配体生成、ScorePosNet3D 反向采样、扩散似然估计、PropPred 性质预测。","典型输入：protein_pos·v / ligand 初始化 / config·ckpt / num_steps·center_pos_mode","典型输出：配体 pos·type / 轨迹 / loss·likelihood / 性质分数","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience TargetDiff 入口（examples/biosciences/targetdiff）与 ScorePosNet3D.sample_diffusion / get_diffusion_loss 是否可用。","3. 如果需要模型权重、数据库、特征缓存或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：口袋条件配体生成、ScorePosNet3D 反向采样、扩散似然估计、PropPred 性质预测。"].join(`
-`)},scenarios:[{id:"targetdiff-gen",title:"口袋条件配体生成",summary:"给定蛋白口袋采样新的三维配体坐标与原子类型",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：靶点口袋 de novo 配体"},{id:"inputData",label:"口袋结构 / ckpt",placeholder:"例如：pocket PDB、featurizer、ScorePosNet3D ckpt、原子数",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成可后处理成键的候选",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"TargetDiff 口袋生成",keyConfig:"protein_atom_feature_dim、ligand_atom_feature_dim、center_pos_mode",resources:"GPU；生成后需化学后处理",deliverables:"ligand pos/type、候选 ID、后处理提示"},requiredData:["biology_targetdiff_dataset 口袋特征","匹配 config/ckpt","配体原子数初始化"],expectedOutputs:"生成配体张量与成键后处理入口",promptTemplate:["请基于 {{system}} 完成 TargetDiff「口袋条件配体生成」。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请用与训练一致的 featurizer 构造 protein_pos/v 与 batch，","实例化 ScorePosNet3D 加载权重，初始化配体后 sample_diffusion；","强调输出需成键/价态后处理才是分子，再交 docking/打分。","如需安装依赖，请先向我请求批准。"].join(`
-`)},{id:"targetdiff-sample",title:"ScorePosNet3D 反向采样",summary:"控制 num_steps/pos_only 等做反向扩散采样与轨迹导出",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：调试采样质量"},{id:"inputData",label:"初始化 / 步数",placeholder:"例如：init ligand_pos/v、num_steps、pos_only、轨迹开关",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：比较步数对构象影响",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"ScorePosNet3D.sample_diffusion",keyConfig:"num_steps、center_pos_mode、pos_only、轨迹",resources:"GPU；减步数可加速但影响质量",deliverables:"最终 pos/type、可选轨迹"},requiredData:["初始化配体张量","ScorePosNet3D ckpt","采样超参"],expectedOutputs:"采样结果与超参对比",promptTemplate:["请基于 {{system}} 运行 ScorePosNet3D 反向采样。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请设置 num_steps/center_pos_mode/pos_only，调用 sample_diffusion，","检查坐标数值与 batch mapping；OOM 时减 batch 或轨迹。"].join(`
-`)},{id:"targetdiff-nll",title:"扩散似然估计",summary:"对给定配体构象估计 diffusion likelihood/KL",fields:[{id:"system",label:"研究对象 / 生物体系",placeholder:"例如：生成物 vs 参考配体"},{id:"inputData",label:"口袋+配体张量",placeholder:"例如：protein/ligand 特征、ckpt",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：比较候选 likelihood",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 生物信息流水线"}],capabilitySummary:{current:"likelihood_estimation",keyConfig:"与训练 schedule 一致、batch 对齐",resources:"GPU",deliverables:"likelihood/KL、排序表"},requiredData:["口袋与配体张量","匹配 diffusion schedule 的 ckpt"],expectedOutputs:"似然分数表",promptTemplate:["请基于 {{system}} 做 TargetDiff 扩散似然估计。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请调用 likelihood_estimation，确保 feature dims/atom vocabulary 与 ckpt 一致，","输出可比较分数；异常时查 center mode 与 batch mapping。"].join(`
-`)},{id:"targetdiff-prop",title:"PropPred 性质预测",summary:"用独立 PropPredNet/Enc 预测生成配体相关性质",fields:t,capabilitySummary:{current:"PropPredNet 性质头",keyConfig:"PropPredNet 或 PropPredNetEnc、与扩散模型分离构造",resources:"GPU/CPU",deliverables:"性质分数、筛选表"},requiredData:["配体 embedding 或结构特征","PropPred 权重","性质定义"],expectedOutputs:"性质预测表与筛选结论",promptTemplate:["请基于 {{system}} 完成 TargetDiff PropPred 性质预测。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请独立实例化 PropPredNet/PropPredNetEnc（可先 fetch_embedding），","输出性质分数用于筛选；无效分子率高时先查 dataset/词表与后处理。"].join(`
-`)}]},O=[l,d,p,m,c,u,_,h,f,y,g,b,D,k,S,P,C,M],v={id:"BE_MPNN",name:"BE-MPNN",category:"onemodel",domain:"fluid",tags:["流体","边界嵌入","异构图","椭圆PDE"],blurb:"BENO 边界嵌入异构图网络，用 G1/G2 图对复杂边界椭圆 PDE 做节点解场代理。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/cfd/BENO",intro:{title:"30 秒看懂模型",subtitle:"用 RHS/SOL/BC 构造异构图，经边界 Transformer 调制消息传递，预测不规则域采样点解场",suited:["边界嵌入椭圆 PDE 代理","RHS/SOL/BC 异构图训练","BENO 边界条件感知推理","复杂边界解场重建评估"],notFor:["无 RHS/SOL/BC 协议","纯规则网格无复杂边界","无法构造 G1/G2 异构图"],prepare:["RHS/SOL/BC npy","BENODatapipe / HeteroData","conf/beno.yaml","边界点数与特征维"],obtain:["节点解场预测","解码后 L2/MAE","checkpoint","重建网格图"],steps:["核对 RHS/SOL/BC","构造 G1/G2 图","配置 BE_MPNN 维数","训练与解码评估","推理回填网格"]},install:{banner:"已为你创建 BE-MPNN 环境配置任务",prompt:["请为 BE-MPNN 创建一个可以直接跑通的环境配置任务。","模型定位：BENO 边界嵌入异构图消息传递模型（HeteroGNS/BE_MPNN）。","适合任务：边界嵌入椭圆 PDE 代理、RHS/SOL/BC 异构图训练、BENO 边界条件感知推理、复杂边界解场重建评估。","典型输入：RHS/SOL/BC npy / BENODatapipe / HeteroData / conf/beno.yaml / 边界点数与特征维。","典型输出：节点解场预测 / 解码后 L2/MAE / checkpoint / 重建网格图。","关键配置：nnode_in_features、nedge_in_features、nnode_out_features、latent_dim、nmessage_passing_steps、boundary_dim、trans_layer。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（examples/cfd/BENO（train.py / inference.py / conf/beno.yaml））与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：边界嵌入椭圆 PDE 代理 / RHS/SOL/BC 异构图训练 / BENO 边界条件感知推理。"].join(`
-`)},scenarios:[{id:"BE_MPNN-boundary-elliptic",title:"边界嵌入椭圆 PDE 代理",summary:"在复杂边界形状与非均匀 BC 下训练椭圆 PDE 节点解场代理",fields:[{...e[0],placeholder:"例如：带不规则边界的 Poisson/椭圆 PDE 域"},{...e[1],placeholder:"例如：RHS_*/SOL_*/BC_*.npy、resolution、file_prefix、cache_dir"},e[2],e[3]],capabilitySummary:{current:"边界嵌入椭圆 PDE 代理",keyConfig:"nnode_in/out_features、nedge_in_features、latent_dim、nmessage_passing_steps、boundary_dim",resources:"远程 GPU；先单卡冒烟 HeteroData forward，再正式训练",deliverables:"checkpoint、解码后 L2/MAE、训练日志与 beno.yaml 快照"},requiredData:["RHS/SOL/BC npy 与 file_prefix","resolution/ns 与节点边规模","目标通道 nnode_out_features"],expectedOutputs:"checkpoint、误差表、配置快照",promptTemplate:["请基于 {{system}} 体系完成 BE-MPNN「边界嵌入椭圆 PDE 代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience BE-MPNN（examples/cfd/BENO（train.py / inference.py / conf/beno.yaml））执行，关键参数关注：nnode_in_features、nedge_in_features、nnode_out_features、latent_dim、nmessage_passing_steps、boundary_dim、trans_layer。","1. 核对 RHS/SOL/BC 字段与边界 (BoundaryNodes,3)。","2. 跑 BENODatapipe 生成 G1/G2/G1+2，检查特征维。","3. 按图协议设 nnode/nedge/latent/message_passing/boundary_dim。","4. 短训冒烟确认 loss 与 u_normalizer.decode。","5. 正式训练并输出 L2/MAE 与 checkpoint 路径。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"BE_MPNN-hetero-train",title:"RHS/SOL/BC 异构图训练",summary:"按 BENO 数据管道把三类数组编成 G1/G2 图并训练",fields:[{...e[0],placeholder:"例如：带不规则边界的 Poisson/椭圆 PDE 域"},{...e[1],placeholder:"例如：RHS_*/SOL_*/BC_*.npy、resolution、file_prefix、cache_dir"},e[2],e[3]],capabilitySummary:{current:"RHS/SOL/BC 异构图训练",keyConfig:"file_prefix、resolution、ntrain/ntest、cache_dir、save_period",resources:"GPU；大分辨率注意边数与 batch",deliverables:"训练曲线、缓存图诊断、最佳 checkpoint"},requiredData:["三类 npy 齐全","与 resolution 匹配的采样","conf/beno.yaml"],expectedOutputs:"训练产物与缓存一致性报告",promptTemplate:["请基于 {{system}} 体系完成 BE-MPNN「RHS/SOL/BC 异构图训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience BE-MPNN（examples/cfd/BENO（train.py / inference.py / conf/beno.yaml））执行，关键参数关注：nnode_in_features、nedge_in_features、nnode_out_features、latent_dim、nmessage_passing_steps、boundary_dim、trans_layer。","1. 校验三类文件与 ntrain/ntest。","2. 预处理并缓存 HeteroData，避免旧 .pt 误用。","3. 对齐 G1.x / edge_features / G1+2.y 顺序。","4. 启动 train.py，按 save_period 存盘。","5. 记录失败时先打印 G1/G2 字段而非改解码器。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"BE_MPNN-bc-infer",title:"BENO 边界条件感知推理",summary:"加载 checkpoint 对测试图做边界调制前向与回填",fields:[{...e[0],placeholder:"例如：带不规则边界的 Poisson/椭圆 PDE 域"},{...e[1],placeholder:"例如：RHS_*/SOL_*/BC_*.npy、resolution、file_prefix、cache_dir"},e[2],e[3]],capabilitySummary:{current:"BENO 边界条件感知推理",keyConfig:"checkpoint、u_normalizer、sample_idx 回填、DDP module. 前缀",resources:"单卡即可；大批次可多卡",deliverables:"预测数组、重建网格图、与真值对比图"},requiredData:["checkpoint","测试 loader / BC 边界序列","normalizer 与训练一致"],expectedOutputs:"预测场、可视化与推理日志",promptTemplate:["请基于 {{system}} 体系完成 BE-MPNN「BENO 边界条件感知推理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience BE-MPNN（examples/cfd/BENO（train.py / inference.py / conf/beno.yaml））执行，关键参数关注：nnode_in_features、nedge_in_features、nnode_out_features、latent_dim、nmessage_passing_steps、boundary_dim、trans_layer。","1. 加载 ckpt（注意 DDP 前缀）。","2. 对 test_loader 前向，保持边界 enc_in=3。","3. 用 sample_idx 回填到 resolution×resolution。","4. decode 到物理尺度并可视化。","5. 异常优先查边界 shape 与缓存配置。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"BE_MPNN-boundary-eval",title:"复杂边界解场重建评估",summary:"在复杂几何附近评估误差并诊断边界编码器容量",fields:[{...e[0],placeholder:"例如：带不规则边界的 Poisson/椭圆 PDE 域"},{...e[1],placeholder:"例如：RHS_*/SOL_*/BC_*.npy、resolution、file_prefix、cache_dir"},e[2],e[3]],capabilitySummary:{current:"复杂边界解场重建评估",keyConfig:"边界采样密度、boundary_dim、trans_layer、几何附近误差切片",resources:"CPU/GPU 均可做评估",deliverables:"分区误差表、边界附近最大误差点、调参建议"},requiredData:["测试真值与预测","边界几何与 sample_idx","训练时 normalizer"],expectedOutputs:"评估报告与下一步调参建议",promptTemplate:["请基于 {{system}} 体系完成 BE-MPNN「复杂边界解场重建评估」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience BE-MPNN（examples/cfd/BENO（train.py / inference.py / conf/beno.yaml））执行，关键参数关注：nnode_in_features、nedge_in_features、nnode_out_features、latent_dim、nmessage_passing_steps、boundary_dim、trans_layer。","1. 按几何分区统计 L2/MAE。","2. 定位边界附近最大误差节点。","3. 检查边界采样与 (BoundaryNodes,3) 协议。","4. 给出增大 boundary_dim/message_passing 或加密边界的建议。","5. 明确是否可进入更长训练或替换主干。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},N={id:"deeponet",name:"DeepONet",category:"onemodel",domain:"fluid",tags:["流体","算子学习","branch-trunk","DeepONet"],blurb:"Branch-Trunk 算子学习：Branch 编码输入函数，Trunk 编码坐标，点积输出物理场。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/DeepONet.py",intro:{title:"30 秒看懂模型",subtitle:"输入函数特征与评估坐标可分离时，用 DeepONet 做稳态/条件算子学习与场回归",suited:["Branch-Trunk 算子拟合","稳态函数到场映射","坐标查询点场回归","CFD Benchmark DeepONet 对比"],notFor:["缺少有效 fx","强依赖局部拓扑","更适合规则网格谱卷积的任务"],prepare:["fx 输入函数","查询坐标 x","fun_dim/out_dim/space_dim","geotype/shapelist"],obtain:["点级物理场","相对 L2","args/YAML","checkpoint"],steps:["核对 fun_dim/out_dim","补齐 branch/trunk_depth","forward 冒烟","训练算子","评估查询点误差"]},install:{banner:"已为你创建 DeepONet 环境配置任务",prompt:["请为 DeepONet 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Branch-Trunk 算子学习模型。","适合任务：Branch-Trunk 算子拟合、稳态函数到场映射、坐标查询点场回归、CFD Benchmark DeepONet 对比。","典型输入：fx 输入函数 / 查询坐标 x / fun_dim/out_dim/space_dim / geotype/shapelist。","典型输出：点级物理场 / 相对 L2 / args/YAML / checkpoint。","关键配置：act、branch_depth、trunk_depth、fun_dim、out_dim、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/DeepONet.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：Branch-Trunk 算子拟合 / 稳态函数到场映射 / 坐标查询点场回归。"].join(`
-`)},scenarios:[{id:"deeponet-branch-trunk",title:"Branch-Trunk 算子拟合",summary:"用 Branch 编码 fx、Trunk 编码坐标，拟合函数到场算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Branch-Trunk 算子拟合",keyConfig:"branch_depth、trunk_depth、n_hidden、fun_dim、out_dim、act",resources:"远程 GPU；点数大时降 batch",deliverables:"checkpoint、训练曲线、查询点相对误差"},requiredData:["含 fx 与坐标的 datapipe","fun_dim/out_dim 协议","geotype/shapelist"],expectedOutputs:"算子 checkpoint 与误差报告",promptTemplate:["请基于 {{system}} 体系完成 DeepONet「Branch-Trunk 算子拟合」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience DeepONet（src/onescience/models/cfd_benchmark/DeepONet.py）执行，关键参数关注：act、branch_depth、trunk_depth、fun_dim、out_dim、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 确认 datapipe 输出 fx 与评估坐标可分离。","2. 按协议设 fun_dim/out_dim/space_dim/geotype。","3. 补齐 branch_depth/trunk_depth/n_hidden，勿猜默认。","4. 假 batch 检查输出 shape==out_dim。","5. 接入训练与相对 L2，输出 checkpoint。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"deeponet-steady-map",title:"稳态函数到场映射",summary:"在稳态任务上把输入函数映射为全场解",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"稳态函数到场映射",keyConfig:"task=steady、unified_pos、ref、归一化",resources:"单卡冒烟后多卡",deliverables:"稳态预测场、误差图、配置快照"},requiredData:["稳态样本与标签","通道维一致的 args"],expectedOutputs:"预测场与评估表",promptTemplate:["请基于 {{system}} 体系完成 DeepONet「稳态函数到场映射」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience DeepONet（src/onescience/models/cfd_benchmark/DeepONet.py）执行，关键参数关注：act、branch_depth、trunk_depth、fun_dim、out_dim、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 设 task 为 steady 并核对归一化。","2. 校验 NumPoints 与 shapelist（结构网格）。","3. 训练至验证误差稳定。","4. 输出场图与相对误差。","5. 说明可否进入动态任务。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"deeponet-query-regress",title:"坐标查询点场回归",summary:"在任意查询坐标上回归物理量，验证 Trunk 泛化",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"坐标查询点场回归",keyConfig:"查询点分布、Trunk 输入、out_dim",resources:"推理可 CPU/GPU",deliverables:"查询点预测、与真值对比、外推说明"},requiredData:["训练好的 DeepONet","查询坐标集","对应 fx"],expectedOutputs:"查询点结果表与外推风险说明",promptTemplate:["请基于 {{system}} 体系完成 DeepONet「坐标查询点场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience DeepONet（src/onescience/models/cfd_benchmark/DeepONet.py）执行，关键参数关注：act、branch_depth、trunk_depth、fun_dim、out_dim、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 加载权重并固定 fx 编码。","2. 在训练域内外采样查询点。","3. 对比误差并标出外推区。","4. 检查 Trunk 坐标尺度是否与训练一致。","5. 给出可用查询范围建议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"deeponet-bench-deeponet",title:"CFD Benchmark DeepONet 对比",summary:"在同一 datapipe 上与 FNO/GNOT 等做算子基线对比",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"CFD Benchmark DeepONet 对比",keyConfig:"相同划分/归一化、fun_dim/out_dim、对比指标",resources:"多模型同卡对比建议固定 seed",deliverables:"对比表、配置差异说明、选型建议"},requiredData:["统一 CFD_Benchmark 协议","各模型可实例化 args"],expectedOutputs:"对比报告与选型结论",promptTemplate:["请基于 {{system}} 体系完成 DeepONet「CFD Benchmark DeepONet 对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience DeepONet（src/onescience/models/cfd_benchmark/DeepONet.py）执行，关键参数关注：act、branch_depth、trunk_depth、fun_dim、out_dim、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 锁定数据划分与归一化。","2. 为 DeepONet 生成完整 args。","3. 同指标评估相对 L2。","4. 记录 DeepONet 相对谱/图模型优缺点。","5. 给出是否继续加深 branch/trunk 的建议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},F={id:"f_fno",name:"F-FNO",category:"onemodel",domain:"fluid",tags:["流体","谱卷积","因子化FNO","规则网格"],blurb:"分解式 Fourier 算子：沿各轴因子化谱卷积，在 FNO 基础上控制多维计算量。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/F_FNO.py",intro:{title:"30 秒看懂模型",subtitle:"规则网格多维场预测时，用因子化谱卷积做高效频域场到场代理",suited:["分解式谱卷积训练","多维网格频域加速对比","Factorized FNO 场预测","modes/shapelist 谱截断调参"],notFor:["点集无法投影到规则网格","shapelist 与真实网格不一致","无结构拓扑优先任务"],prepare:["规则网格场","fun_dim/out_dim","modes","geotype/shapelist"],obtain:["预测场","谱截断误差","checkpoint","对比基线 FNO 的耗时"],steps:["核对 shapelist","设 modes/n_layers","谱卷积冒烟","训练场代理","对比 FNO 算力"]},install:{banner:"已为你创建 F-FNO 环境配置任务",prompt:["请为 F-FNO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark 因子化 Fourier Neural Operator。","适合任务：分解式谱卷积训练、多维网格频域加速对比、Factorized FNO 场预测、modes/shapelist 谱截断调参。","典型输入：规则网格场 / fun_dim/out_dim / modes / geotype/shapelist。","典型输出：预测场 / 谱截断误差 / checkpoint / 对比基线 FNO 的耗时。","关键配置：act、fun_dim、out_dim、modes、n_hidden、n_layers、geotype、shapelist、space_dim、time_input、unified_pos、ref。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/F_FNO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：分解式谱卷积训练 / 多维网格频域加速对比 / Factorized FNO 场预测。"].join(`
-`)},scenarios:[{id:"f_fno-fact-spectral",title:"分解式谱卷积训练",summary:"沿空间轴分解 Fourier 卷积训练多维场代理",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"分解式谱卷积训练",keyConfig:"modes、n_layers、n_hidden、fun_dim、out_dim、shapelist",resources:"远程 GPU；3D 时优先降 modes/batch",deliverables:"checkpoint、训练曲线、相对 L2"},requiredData:["可 reshape 的规则网格","与 shapelist 一致的点数","通道协议"],expectedOutputs:"训练权重与误差",promptTemplate:["请基于 {{system}} 体系完成 F-FNO「分解式谱卷积训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience F-FNO（src/onescience/models/cfd_benchmark/F_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、n_layers、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 验证 NumPoints==prod(shapelist)。","2. 按数据设 fun_dim/out_dim/geotype。","3. 配置 modes/n_layers 并假 batch 检查输出。","4. 短训确认谱卷积稳定。","5. 正式训练输出相对误差与 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"f_fno-ndim-speed",title:"多维网格频域加速对比",summary:"与标准 FNO 对比同精度下的显存与步时",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"多维网格频域加速对比",keyConfig:"同 modes/数据、计时、显存峰值",resources:"同卡同 batch 对比",deliverables:"耗时/显存对比表、精度并列结果"},requiredData:["FNO 与 F_FNO 可跑配置","同一测试集"],expectedOutputs:"加速结论与选型建议",promptTemplate:["请基于 {{system}} 体系完成 F-FNO「多维网格频域加速对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience F-FNO（src/onescience/models/cfd_benchmark/F_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、n_layers、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 固定数据与 modes。","2. 分别计时 forward/train step。","3. 对比相对 L2。","4. 记录显存峰值。","5. 给出何时选因子化路线。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"f_fno-ffno-predict",title:"Factorized FNO 场预测",summary:"加载权重对规则网格新样本做场到场推理",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Factorized FNO 场预测",keyConfig:"checkpoint、geotype、shapelist、归一化",resources:"单卡推理",deliverables:"预测场、误差图、失败样本列表"},requiredData:["权重","与训练一致的网格协议"],expectedOutputs:"预测结果与诊断",promptTemplate:["请基于 {{system}} 体系完成 F-FNO「Factorized FNO 场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience F-FNO（src/onescience/models/cfd_benchmark/F_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、n_layers、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 加载 ckpt 核对 modes/shapelist。","2. 推理并反归一化。","3. 计算测试相对误差。","4. 检查 reshape 失败样本。","5. 必要时建议投影 adapter。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"f_fno-modes-tune",title:"modes/shapelist 谱截断调参",summary:"扫描 modes 与网格尺寸，平衡精度与算力",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"modes/shapelist 谱截断调参",keyConfig:"modes 扫描、shapelist、n_layers",resources:"短训网格搜索即可",deliverables:"modes-误差曲线、推荐配置"},requiredData:["验证集","可改的 YAML/args"],expectedOutputs:"推荐 modes 与风险说明",promptTemplate:["请基于 {{system}} 体系完成 F-FNO「modes/shapelist 谱截断调参」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience F-FNO（src/onescience/models/cfd_benchmark/F_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、n_layers、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 在小预算下扫描 modes。","2. 确认 shapelist 未被错误 padding。","3. 记录误差与步时。","4. 选出帕累托点。","5. 写入最终配置并冒烟。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},T={id:"factformer",name:"Factformer",category:"onemodel",domain:"fluid",tags:["流体","因子化注意力","结构化网格","Transformer"],blurb:"因子化注意力 Transformer：在结构化网格上用更省的全局交互做场预测。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/factformer",intro:{title:"30 秒看懂模型",subtitle:"结构化中大规模网格场预测，用 Factformer_block 替代稠密自注意力做全局建模",suited:["因子化注意力场建模","结构化网格全局交互训练","Factformer token 场回归","大网格省显存注意力对比"],notFor:["geotype=unstructured（源码会抛错）","非结构点云无投影","极小网格无需注意力"],prepare:["structured 网格","坐标/位置编码","fun_dim/out_dim","n_heads/n_layers/shapelist"],obtain:["token 场预测","相对误差","checkpoint","显存对比日志"],steps:["确认非 unstructured","设 n_heads/mlp_ratio","token 冒烟","训练","对比稠密注意力"]},install:{banner:"已为你创建 Factformer 环境配置任务",prompt:["请为 Factformer 创建一个可以直接跑通的环境配置任务。","模型定位：因子化注意力场 Transformer（onescience.models.factformer / CFD_Benchmark）。","适合任务：因子化注意力场建模、结构化网格全局交互训练、Factformer token 场回归、大网格省显存注意力对比。","典型输入：structured 网格 / 坐标/位置编码 / fun_dim/out_dim / n_heads/n_layers/shapelist。","典型输出：token 场预测 / 相对误差 / checkpoint / 显存对比日志。","关键配置：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/factformer）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：因子化注意力场建模 / 结构化网格全局交互训练 / Factformer token 场回归。"].join(`
-`)},scenarios:[{id:"factformer-fact-attn",title:"因子化注意力场建模",summary:"用 Factformer_block 在结构网格上建模全局场依赖",fields:[{...e[0],placeholder:"例如：结构化 2D/3D CFD 场、中等以上分辨率"},{...e[1],placeholder:"例如：structured geotype、shapelist、fun_dim/out_dim、n_heads"},e[2],e[3]],capabilitySummary:{current:"因子化注意力场建模",keyConfig:"n_heads、n_layers、mlp_ratio、dropout、n_hidden",resources:"远程 GPU；分辨率高时降 batch",deliverables:"checkpoint、训练曲线、相对 L2"},requiredData:["geotype≠unstructured","shapelist 与点数一致","通道协议"],expectedOutputs:"权重与误差",promptTemplate:["请基于 {{system}} 体系完成 Factformer「因子化注意力场建模」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Factformer（src/onescience/models/factformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 拒绝 unstructured，确认 structured_*。","2. 补齐注意力相关 args。","3. 拼接位置编码后做 forward shape 测试。","4. 短训确认稳定。","5. 正式训练并保存最佳 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"factformer-struct-global",title:"结构化网格全局交互训练",summary:"在中大规模结构网格上训练全局交互场代理",fields:[{...e[0],placeholder:"例如：结构化 2D/3D CFD 场、中等以上分辨率"},{...e[1],placeholder:"例如：structured geotype、shapelist、fun_dim/out_dim、n_heads"},e[2],e[3]],capabilitySummary:{current:"结构化网格全局交互训练",keyConfig:"shapelist、unified_pos、time_input、batch",resources:"多卡可选；先单卡窗口",deliverables:"训练日志、显存曲线、验证误差"},requiredData:["结构网格数据","完整 Factformer args"],expectedOutputs:"可复现训练包",promptTemplate:["请基于 {{system}} 体系完成 Factformer「结构化网格全局交互训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Factformer（src/onescience/models/factformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 验证 NumPoints==prod(shapelist)。","2. 配置窗口/batch 防 OOM。","3. 训练并监控验证误差。","4. 记录显存峰值。","5. 给出扩分辨率计划。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"factformer-token-regress",title:"Factformer token 场回归",summary:"将坐标与场特征映射为 token 后回归目标场",fields:[{...e[0],placeholder:"例如：结构化 2D/3D CFD 场、中等以上分辨率"},{...e[1],placeholder:"例如：structured geotype、shapelist、fun_dim/out_dim、n_heads"},e[2],e[3]],capabilitySummary:{current:"Factformer token 场回归",keyConfig:"fun_dim、out_dim、ref、act",resources:"GPU 推理/训练",deliverables:"预测场、token 维度诊断、误差图"},requiredData:["(x,fx) 协议数据","out_dim 标签"],expectedOutputs:"回归结果与诊断",promptTemplate:["请基于 {{system}} 体系完成 Factformer「Factformer token 场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Factformer（src/onescience/models/factformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 检查 token 维与 n_hidden。","2. 冒烟输出通道。","3. 训练/微调。","4. 可视化误差。","5. 异常先查位置编码尺度。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"factformer-save-mem-attn",title:"大网格省显存注意力对比",summary:"与标准 Transformer 对比同网格下的显存与精度",fields:[{...e[0],placeholder:"例如：结构化 2D/3D CFD 场、中等以上分辨率"},{...e[1],placeholder:"例如：structured geotype、shapelist、fun_dim/out_dim、n_heads"},e[2],e[3]],capabilitySummary:{current:"大网格省显存注意力对比",keyConfig:"同 n_layers/数据、显存、相对 L2",resources:"同卡对比",deliverables:"显存-精度对比表、选型建议"},requiredData:["Transformer 与 Factformer 配置","同一网格"],expectedOutputs:"对比报告",promptTemplate:["请基于 {{system}} 体系完成 Factformer「大网格省显存注意力对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Factformer（src/onescience/models/factformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 固定网格与训练步数。","2. 分别测显存与误差。","3. 记录加速比。","4. 标出 Factformer 适用分辨率区间。","5. 给出是否换谱算子的备选。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},A={id:"fno",name:"FNO",category:"onemodel",domain:"fluid",tags:["流体","FNO","谱卷积","神经算子"],blurb:"Fourier Neural Operator：规则网格上谱卷积做全局低频场到场建模的强基线。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/FNO.py",intro:{title:"30 秒看懂模型",subtitle:"规则网格稳态/动态场预测，用 MLP 升维 + 1D/2D/3D 谱卷积学习全局算子",suited:["Fourier 神经算子基线训练","规则网格低频场代理","FNO 稳态/动态场预测","Geo 投影网格谱卷积推理"],notFor:["无法 reshape 为 shapelist 的点云","极强高频局部细节不适谱截断","无网格协议"],prepare:["规则网格或 Geo 投影","fun_dim/out_dim","modes","geotype/shapelist"],obtain:["预测场","相对 L2","checkpoint","谱截断诊断"],steps:["核对 shapelist/modes","实例化 FNO","forward 冒烟","训练","评估低频误差"]},install:{banner:"已为你创建 FNO 环境配置任务",prompt:["请为 FNO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Fourier Neural Operator。","适合任务：Fourier 神经算子基线训练、规则网格低频场代理、FNO 稳态/动态场预测、Geo 投影网格谱卷积推理。","典型输入：规则网格或 Geo 投影 / fun_dim/out_dim / modes / geotype/shapelist。","典型输出：预测场 / 相对 L2 / checkpoint / 谱截断诊断。","关键配置：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/FNO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：Fourier 神经算子基线训练 / 规则网格低频场代理 / FNO 稳态/动态场预测。"].join(`
-`)},scenarios:[{id:"fno-fno-baseline",title:"Fourier 神经算子基线训练",summary:"配置 modes 与通道，在 CFD_Benchmark 上训练 FNO 基线",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Fourier 神经算子基线训练",keyConfig:"modes、n_hidden、fun_dim、out_dim、geotype、shapelist",resources:"远程 GPU；先 dry-run args",deliverables:"checkpoint、曲线、相对 L2"},requiredData:["可谱卷积的网格数据","完整 args"],expectedOutputs:"基线权重与指标",promptTemplate:["请基于 {{system}} 体系完成 FNO「Fourier 神经算子基线训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FNO（src/onescience/models/cfd_benchmark/FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 确认 geotype 与 shapelist。","2. 设 modes 与通道维。","3. 假输入检查输出 out_dim。","4. 短训冒烟。","5. 正式训练保存 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fno-lowfreq-surr",title:"规则网格低频场代理",summary:"强调全局低频结构的场代理，适合 Darcy 类问题",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"规则网格低频场代理",keyConfig:"modes 截断、归一化、相对 L2",resources:"单卡即可中小网格",deliverables:"代理场、频谱误差分析、可用性结论"},requiredData:["规则网格标签","训练配置"],expectedOutputs:"代理结果与频谱说明",promptTemplate:["请基于 {{system}} 体系完成 FNO「规则网格低频场代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FNO（src/onescience/models/cfd_benchmark/FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 训练/加载 FNO。","2. 在测试集评估。","3. 分析残余高频误差。","4. 判断是否需提高 modes 或换 U_FNO。","5. 给出部署建议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fno-fno-dyn",title:"FNO 稳态/动态场预测",summary:"按 task 配置稳态或 time_input 动态条件预测",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"FNO 稳态/动态场预测",keyConfig:"time_input、task、T 通道、shapelist",resources:"动态序列建议 GPU",deliverables:"稳态或时序预测、逐步误差"},requiredData:["含时间协议或稳态样本","匹配的 args"],expectedOutputs:"预测序列/场与误差",promptTemplate:["请基于 {{system}} 体系完成 FNO「FNO 稳态/动态场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FNO（src/onescience/models/cfd_benchmark/FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 明确 steady 或 dynamic_*。","2. 对齐 time_input 与 datapipe。","3. 训练或推理。","4. 输出逐步/全场误差。","5. 滚动任务警惕误差累积。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fno-geo-spectral",title:"Geo 投影网格谱卷积推理",summary:"非结构点经 Geo 投影后用 FNO 谱卷积推理",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Geo 投影网格谱卷积推理",keyConfig:"unified_pos、Geo 投影、ref、shapelist",resources:"投影+推理 GPU",deliverables:"投影场、反投影误差、风险说明"},requiredData:["投影 adapter 或 Geo 流程","FNO 权重"],expectedOutputs:"推理结果与投影质量报告",promptTemplate:["请基于 {{system}} 体系完成 FNO「Geo 投影网格谱卷积推理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FNO（src/onescience/models/cfd_benchmark/FNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 检查投影网格与 shapelist。","2. 加载 FNO 推理。","3. 反投影到原点云。","4. 评估投影引入误差。","5. 误差大则建议改 Transolver/图模型。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},x={id:"galerkin_transformer",name:"Galerkin Transformer",category:"onemodel",domain:"fluid",tags:["流体","Galerkin注意力","Transformer","算子学习"],blurb:"Galerkin 风格注意力 Transformer：用近似全局交互做结构化/点级算子基线。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/Galerkin_Transformer.py",intro:{title:"30 秒看懂模型",subtitle:"需要全局 token 交互但希望用 Galerkin attention 近似时的场回归主干",suited:["Galerkin 注意力算子学习","全局 token 交互场回归","Galerkin Transformer 基线对比","点数可控注意力冒烟"],notFor:["点数极大无降采样","注意力显存不可接受","更适合谱卷积的规则网格"],prepare:["token 化场特征","n_heads/n_layers","fun_dim/out_dim","geotype/shapelist"],obtain:["场预测","注意力显存日志","checkpoint","对比指标"],steps:["估 token 数","设 n_heads/mlp_ratio","冒烟","训练","对比稠密注意力"]},install:{banner:"已为你创建 Galerkin Transformer 环境配置任务",prompt:["请为 Galerkin Transformer 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Galerkin Attention Transformer。","适合任务：Galerkin 注意力算子学习、全局 token 交互场回归、Galerkin Transformer 基线对比、点数可控注意力冒烟。","典型输入：token 化场特征 / n_heads/n_layers / fun_dim/out_dim / geotype/shapelist。","典型输出：场预测 / 注意力显存日志 / checkpoint / 对比指标。","关键配置：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/Galerkin_Transformer.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：Galerkin 注意力算子学习 / 全局 token 交互场回归 / Galerkin Transformer 基线对比。"].join(`
-`)},scenarios:[{id:"galerkin_transformer-galerkin-op",title:"Galerkin 注意力算子学习",summary:"用 Galerkin 风格注意力 block 学习全局场算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Galerkin 注意力算子学习",keyConfig:"n_heads、n_layers、mlp_ratio、dropout、n_hidden",resources:"远程 GPU；控制 NumPoints",deliverables:"checkpoint、相对 L2、训练日志"},requiredData:["可 token 化输入","完整 args"],expectedOutputs:"权重与指标",promptTemplate:["请基于 {{system}} 体系完成 Galerkin Transformer「Galerkin 注意力算子学习」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Galerkin Transformer（src/onescience/models/cfd_benchmark/Galerkin_Transformer.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 确认点数预算。","2. 补齐注意力 args。","3. forward 检查 out_dim。","4. 短训。","5. 正式训练存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"galerkin_transformer-token-global",title:"全局 token 交互场回归",summary:"将场编码为 token 后做全局交互回归目标变量",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"全局 token 交互场回归",keyConfig:"fun_dim、out_dim、unified_pos、geotype",resources:"GPU",deliverables:"预测场、误差图"},requiredData:["(x,fx) 数据","标签通道"],expectedOutputs:"回归结果",promptTemplate:["请基于 {{system}} 体系完成 Galerkin Transformer「全局 token 交互场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Galerkin Transformer（src/onescience/models/cfd_benchmark/Galerkin_Transformer.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 编码特征为 token。","2. 训练/推理。","3. 评估相对误差。","4. OOM 时降点数或 layers。","5. 记录是否需换谱模型。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"galerkin_transformer-galerkin-bench",title:"Galerkin Transformer 基线对比",summary:"与标准 Transformer/FNO 在同协议下对比",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"Galerkin Transformer 基线对比",keyConfig:"同划分、同指标、显存对比",resources:"同卡对比",deliverables:"对比表、选型建议"},requiredData:["统一 datapipe","三模型配置"],expectedOutputs:"对比报告",promptTemplate:["请基于 {{system}} 体系完成 Galerkin Transformer「Galerkin Transformer 基线对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Galerkin Transformer（src/onescience/models/cfd_benchmark/Galerkin_Transformer.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 锁定数据与归一化。","2. 跑三模型短评估。","3. 汇总精度/显存。","4. 标出 Galerkin 近似收益。","5. 给出推荐主干。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"galerkin_transformer-token-smoke",title:"点数可控注意力冒烟",summary:"在可控 NumPoints 上验证注意力 mask 与 shape",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"点数可控注意力冒烟",keyConfig:"NumPoints、n_heads、batch、假输入",resources:"单卡冒烟",deliverables:"shape 报告、最大可跑点数估计"},requiredData:["args 草稿","假 batch 生成脚本"],expectedOutputs:"冒烟通过证明与扩容上限",promptTemplate:["请基于 {{system}} 体系完成 Galerkin Transformer「点数可控注意力冒烟」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Galerkin Transformer（src/onescience/models/cfd_benchmark/Galerkin_Transformer.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 构造递增点数假输入。","2. 记录显存与是否报错。","3. 确认输出通道。","4. 写入安全点数上限。","5. 再进入正式训练。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},G={id:"gfno",name:"GFNO",category:"onemodel",domain:"fluid",tags:["流体","等变算子","FNO","对称先验"],blurb:"群等变 Fourier 算子：在 2D 场上引入旋转/反射等变谱卷积与等变 MLP。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/GFNO.py",intro:{title:"30 秒看懂模型",subtitle:"具有明显几何对称性的 2D CFD 场，用等变先验比较相对普通 FNO 的收益",suited:["群等变谱算子训练","2D 对称先验场预测","旋转/反射等变对比实验","GFNO 对称性收益评估"],notFor:["非 2D 场","无对称先验","space_dim/shapelist 与 GFNO 不匹配"],prepare:["2D 规则场","对称性假设说明","modes","fun_dim/out_dim/shapelist"],obtain:["等变预测场","对称增强对比表","checkpoint"],steps:["确认 2D","设 modes","等变冒烟","训练","对称性消融"]},install:{banner:"已为你创建 GFNO 环境配置任务",prompt:["请为 GFNO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Group-Equivariant FNO。","适合任务：群等变谱算子训练、2D 对称先验场预测、旋转/反射等变对比实验、GFNO 对称性收益评估。","典型输入：2D 规则场 / 对称性假设说明 / modes / fun_dim/out_dim/shapelist。","典型输出：等变预测场 / 对称增强对比表 / checkpoint。","关键配置：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/GFNO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：群等变谱算子训练 / 2D 对称先验场预测 / 旋转/反射等变对比实验。"].join(`
-`)},scenarios:[{id:"gfno-equiv-spectral",title:"群等变谱算子训练",summary:"训练带群等变谱卷积的 GFNO 场代理",fields:[{...e[0],placeholder:"例如：具有旋转/反射对称的 2D 流场"},{...e[1],placeholder:"例如：2D 网格、modes、fun_dim/out_dim、对称增强设置"},e[2],e[3]],capabilitySummary:{current:"群等变谱算子训练",keyConfig:"modes、n_hidden、fun_dim、out_dim、2D shapelist",resources:"远程 GPU",deliverables:"checkpoint、曲线、相对 L2"},requiredData:["2D 网格数据","匹配 space_dim"],expectedOutputs:"等变模型权重",promptTemplate:["请基于 {{system}} 体系完成 GFNO「群等变谱算子训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GFNO（src/onescience/models/cfd_benchmark/GFNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 确认 2D geotype/shapelist。","2. 配置 modes 与通道。","3. forward 冒烟。","4. 训练。","5. 保存最佳 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gfno-sym2d-pred",title:"2D 对称先验场预测",summary:"在对称先验场景上推理并检查等变一致性",fields:[{...e[0],placeholder:"例如：具有旋转/反射对称的 2D 流场"},{...e[1],placeholder:"例如：2D 网格、modes、fun_dim/out_dim、对称增强设置"},e[2],e[3]],capabilitySummary:{current:"2D 对称先验场预测",keyConfig:"旋转/反射测试、归一化、modes",resources:"单卡",deliverables:"预测场、等变一致性误差"},requiredData:["权重","可做对称变换的测试样本"],expectedOutputs:"预测与一致性报告",promptTemplate:["请基于 {{system}} 体系完成 GFNO「2D 对称先验场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GFNO（src/onescience/models/cfd_benchmark/GFNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 加载 GFNO。","2. 对样本做群变换前后预测。","3. 度量输出是否近似等变。","4. 评估物理误差。","5. 不一致则查数据增强与通道定义。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gfno-rot-reflect",title:"旋转/反射等变对比实验",summary:"对比 GFNO 与普通 FNO 在对称增强集上的表现",fields:[{...e[0],placeholder:"例如：具有旋转/反射对称的 2D 流场"},{...e[1],placeholder:"例如：2D 网格、modes、fun_dim/out_dim、对称增强设置"},e[2],e[3]],capabilitySummary:{current:"旋转/反射等变对比实验",keyConfig:"同数据同 modes、增强策略、指标",resources:"同卡对比",deliverables:"对比表、对称场景增益"},requiredData:["FNO/GFNO 配置","对称测试集"],expectedOutputs:"对比结论",promptTemplate:["请基于 {{system}} 体系完成 GFNO「旋转/反射等变对比实验」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GFNO（src/onescience/models/cfd_benchmark/GFNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 固定划分与 modes。","2. 分别训练/评估。","3. 在旋转/反射测试上对比。","4. 记录参数量与步时。","5. 给出是否采用等变路线。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gfno-sym-gain",title:"GFNO 对称性收益评估",summary:"量化等变先验带来的样本效率与泛化收益",fields:[{...e[0],placeholder:"例如：具有旋转/反射对称的 2D 流场"},{...e[1],placeholder:"例如：2D 网格、modes、fun_dim/out_dim、对称增强设置"},e[2],e[3]],capabilitySummary:{current:"GFNO 对称性收益评估",keyConfig:"少样本曲线、对称测试误差",resources:"多次短训",deliverables:"样本效率曲线、收益结论"},requiredData:["可降采样训练集","对称测试协议"],expectedOutputs:"收益评估报告",promptTemplate:["请基于 {{system}} 体系完成 GFNO「GFNO 对称性收益评估」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GFNO（src/onescience/models/cfd_benchmark/GFNO.py）执行，关键参数关注：act、fun_dim、out_dim、modes、n_hidden、geotype、shapelist、space_dim、time_input、unified_pos、ref。","1. 在递减训练集上训练 GFNO/FNO。","2. 画误差曲线。","3. 统计对称测试增益。","4. 判断数据是否真正有对称性。","5. 无收益则回退 FNO。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},U={id:"gnot",name:"GNOT",category:"onemodel",domain:"fluid",tags:["流体","神经算子","交叉注意力","双分支"],blurb:"坐标与场特征双分支交互的神经算子 Transformer，适合点级/网格级算子任务。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/GNOT.py",intro:{title:"30 秒看懂模型",subtitle:"位置分支与函数值分支经 GNOT block 交叉建模，学习坐标-场耦合算子",suited:["坐标-场双分支交互训练","GNOT 交叉注意力算子拟合","点级位置函数耦合预测","双分支 fx 信息量诊断"],notFor:["只有单路输入且无有效 fx","双分支优势无法体现","缺坐标分支"],prepare:["坐标 x","场特征 fx","n_heads/n_layers","fun_dim/out_dim"],obtain:["耦合场预测","分支消融结果","checkpoint"],steps:["检查双分支输入","设注意力超参","冒烟","训练","fx 信息量诊断"]},install:{banner:"已为你创建 GNOT 环境配置任务",prompt:["请为 GNOT 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Graph Neural Operator Transformer (GNOT)。","适合任务：坐标-场双分支交互训练、GNOT 交叉注意力算子拟合、点级位置函数耦合预测、双分支 fx 信息量诊断。","典型输入：坐标 x / 场特征 fx / n_heads/n_layers / fun_dim/out_dim。","典型输出：耦合场预测 / 分支消融结果 / checkpoint。","关键配置：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/GNOT.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：坐标-场双分支交互训练 / GNOT 交叉注意力算子拟合 / 点级位置函数耦合预测。"].join(`
-`)},scenarios:[{id:"gnot-coord-field",title:"坐标-场双分支交互训练",summary:"分别预处理坐标与场特征后经 GNOT block 训练",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"坐标-场双分支交互训练",keyConfig:"n_heads、n_layers、mlp_ratio、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、曲线、相对 L2"},requiredData:["同时含 x 与有信息 fx","完整 args"],expectedOutputs:"双分支模型权重",promptTemplate:["请基于 {{system}} 体系完成 GNOT「坐标-场双分支交互训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GNOT（src/onescience/models/cfd_benchmark/GNOT.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 确认 fx 非退化。","2. 配置双分支与注意力超参。","3. shape 冒烟。","4. 训练。","5. 保存 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gnot-gnot-cross",title:"GNOT 交叉注意力算子拟合",summary:"用交叉注意力拟合位置与函数值之间的算子映射",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"GNOT 交叉注意力算子拟合",keyConfig:"dropout、n_hidden、geotype、unified_pos",resources:"GPU",deliverables:"算子拟合结果、注意力诊断"},requiredData:["datapipe 双分支输出","标签"],expectedOutputs:"拟合报告",promptTemplate:["请基于 {{system}} 体系完成 GNOT「GNOT 交叉注意力算子拟合」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GNOT（src/onescience/models/cfd_benchmark/GNOT.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 对齐 datapipe 双分支 schema。","2. 训练交叉注意力块。","3. 评估查询点/网格误差。","4. OOM 降点数。","5. 记录与 DeepONet/Transformer 差异。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gnot-pos-couple",title:"点级位置函数耦合预测",summary:"在点级任务上预测与位置强耦合的物理场",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"点级位置函数耦合预测",keyConfig:"space_dim、fun_dim、out_dim、ref",resources:"单卡/多卡",deliverables:"点级预测、误差分布"},requiredData:["点云或网格节点数据"],expectedOutputs:"预测与误差",promptTemplate:["请基于 {{system}} 体系完成 GNOT「点级位置函数耦合预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GNOT（src/onescience/models/cfd_benchmark/GNOT.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 检查坐标尺度。","2. 训练/推理。","3. 按区域统计误差。","4. 弱耦合时检查 fx 设计。","5. 给出特征工程建议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"gnot-fx-diag",title:"双分支 fx 信息量诊断",summary:"诊断 fx 是否提供有效信息，避免双分支退化",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"双分支 fx 信息量诊断",keyConfig:"fx 消融、置换检验、通道贡献",resources:"短评估即可",deliverables:"信息量报告、是否改单分支建议"},requiredData:["可消融的 fx","验证集"],expectedOutputs:"诊断结论",promptTemplate:["请基于 {{system}} 体系完成 GNOT「双分支 fx 信息量诊断」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GNOT（src/onescience/models/cfd_benchmark/GNOT.py）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 做 fx 置零/打乱对照。","2. 对比误差变化。","3. 检查通道相关。","4. 若无增益建议换模型或重做特征。","5. 有增益则固化特征协议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},E={id:"graph_unet",name:"Graph U-Net",category:"onemodel",domain:"fluid",tags:["流体","图U-Net","多尺度","非结构网格"],blurb:"图上的 U-Net：GCN+TopK 池化/上采样，对非结构网格做多尺度节点场回归。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/Graph_UNet.py",intro:{title:"30 秒看懂模型",subtitle:"图拓扑稳定且有多尺度局部结构时，用编码-解码图 U-Net 预测节点物理场",suited:["图 U-Net 多尺度节点回归","TopK 池化编码解码训练","非结构多尺度场重建","图索引映射冒烟检查"],notFor:["点序/边不稳定","池化索引无法维护","无图结构"],prepare:["edge_index/图","fun_dim/out_dim","n_hidden","稳定点序"],obtain:["节点场","多尺度诊断","checkpoint"],steps:["检查图连通","设 n_hidden","池化冒烟","训练","校验索引映射"]},install:{banner:"已为你创建 Graph U-Net 环境配置任务",prompt:["请为 Graph U-Net 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Graph U-Net。","适合任务：图 U-Net 多尺度节点回归、TopK 池化编码解码训练、非结构多尺度场重建、图索引映射冒烟检查。","典型输入：edge_index/图 / fun_dim/out_dim / n_hidden / 稳定点序。","典型输出：节点场 / 多尺度诊断 / checkpoint。","关键配置：act、fun_dim、out_dim、n_hidden。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/Graph_UNet.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：图 U-Net 多尺度节点回归 / TopK 池化编码解码训练 / 非结构多尺度场重建。"].join(`
-`)},scenarios:[{id:"graph_unet-gunet-multi",title:"图 U-Net 多尺度节点回归",summary:"在图 U-Net 上训练多尺度节点物理场回归",fields:[{...e[0],placeholder:"例如：非结构网格多尺度节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_hidden"},e[2],e[3]],capabilitySummary:{current:"图 U-Net 多尺度节点回归",keyConfig:"n_hidden、fun_dim、out_dim、act",resources:"远程 GPU",deliverables:"checkpoint、相对误差、训练日志"},requiredData:["稳定图","节点标签"],expectedOutputs:"权重与指标",promptTemplate:["请基于 {{system}} 体系完成 Graph U-Net「图 U-Net 多尺度节点回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Graph U-Net（src/onescience/models/cfd_benchmark/Graph_UNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden。","1. 确认 edge_index 稳定。","2. 配置通道与 n_hidden。","3. forward 冒烟。","4. 训练。","5. 评估节点误差。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graph_unet-topk-codec",title:"TopK 池化编码解码训练",summary:"用 TopKPooling 与上采样路径做图编码解码训练",fields:[{...e[0],placeholder:"例如：非结构网格多尺度节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_hidden"},e[2],e[3]],capabilitySummary:{current:"TopK 池化编码解码训练",keyConfig:"池化比例、层数、索引回映",resources:"GPU；注意池化后节点数",deliverables:"训练曲线、池化节点数日志"},requiredData:["可池化的连通图"],expectedOutputs:"训练产物",promptTemplate:["请基于 {{system}} 体系完成 Graph U-Net「TopK 池化编码解码训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Graph U-Net（src/onescience/models/cfd_benchmark/Graph_UNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden。","1. 记录各层池化后节点数。","2. 确保上采样索引可逆。","3. 短训观察梯度。","4. 正式训练。","5. 失败先查索引而非 loss。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graph_unet-unstruct-recon",title:"非结构多尺度场重建",summary:"对非结构网格目标场做多尺度重建与可视化",fields:[{...e[0],placeholder:"例如：非结构网格多尺度节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_hidden"},e[2],e[3]],capabilitySummary:{current:"非结构多尺度场重建",keyConfig:"out_dim、反归一化、可视化映射",resources:"单卡",deliverables:"重建场、误差分布"},requiredData:["权重","测试图"],expectedOutputs:"重建结果",promptTemplate:["请基于 {{system}} 体系完成 Graph U-Net「非结构多尺度场重建」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Graph U-Net（src/onescience/models/cfd_benchmark/Graph_UNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden。","1. 推理。","2. 映射回原点序。","3. 分区误差。","4. 细尺度差则加深通道。","5. 给出是否换 MeshGraphNet。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graph_unet-index-smoke",title:"图索引映射冒烟检查",summary:"在假图上验证池化/解池化索引与输出对齐",fields:[{...e[0],placeholder:"例如：非结构网格多尺度节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_hidden"},e[2],e[3]],capabilitySummary:{current:"图索引映射冒烟检查",keyConfig:"假 edge_index、点序、out_dim",resources:"CPU/GPU 冒烟",deliverables:"shape/索引报告、风险清单"},requiredData:["最小图样例"],expectedOutputs:"冒烟通过证明",promptTemplate:["请基于 {{system}} 体系完成 Graph U-Net「图索引映射冒烟检查」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Graph U-Net（src/onescience/models/cfd_benchmark/Graph_UNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden。","1. 构造小图。","2. 逐步打印池化索引。","3. 确认输出节点数回到输入。","4. 写入约束。","5. 再接入真实数据。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},w={id:"graphsage",name:"GraphSAGE",category:"onemodel",domain:"fluid",tags:["流体","图神经网络","邻域聚合","轻量基线"],blurb:"GraphSAGE 轻量图基线：在 edge_index 上 SAGEConv 邻域聚合，做非结构节点回归。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/GraphSAGE.py",intro:{title:"30 秒看懂模型",subtitle:"已有稳定图拓扑时，用轻量 SAGEConv 多层聚合快速建立非结构场回归下限",suited:["SAGEConv 邻域聚合基线","轻量图拓扑节点回归","edge_index 构图场预测","GraphSAGE CFD 下限对比"],notFor:["无 edge_index","需要复杂多尺度池化","只要点云无图"],prepare:["edge_index","坐标+节点特征","fun_dim/out_dim/space_dim","n_layers/n_hidden"],obtain:["节点预测","相对误差","轻量基线指标"],steps:["检查边","设 n_layers","冒烟","训练","作下限对比"]},install:{banner:"已为你创建 GraphSAGE 环境配置任务",prompt:["请为 GraphSAGE 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark GraphSAGE。","适合任务：SAGEConv 邻域聚合基线、轻量图拓扑节点回归、edge_index 构图场预测、GraphSAGE CFD 下限对比。","典型输入：edge_index / 坐标+节点特征 / fun_dim/out_dim/space_dim / n_layers/n_hidden。","典型输出：节点预测 / 相对误差 / 轻量基线指标。","关键配置：act、fun_dim、out_dim、n_hidden、n_layers、space_dim。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/GraphSAGE.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：SAGEConv 邻域聚合基线 / 轻量图拓扑节点回归 / edge_index 构图场预测。"].join(`
-`)},scenarios:[{id:"graphsage-sage-agg",title:"SAGEConv 邻域聚合基线",summary:"多层 SAGEConv 聚合邻域特征训练节点场基线",fields:[{...e[0],placeholder:"例如：已有网格邻接的非结构 CFD 节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_layers"},e[2],e[3]],capabilitySummary:{current:"SAGEConv 邻域聚合基线",keyConfig:"n_layers、n_hidden、fun_dim、out_dim、act",resources:"单卡即可多数案例",deliverables:"checkpoint、相对 L2"},requiredData:["edge_index","节点标签"],expectedOutputs:"基线权重",promptTemplate:["请基于 {{system}} 体系完成 GraphSAGE「SAGEConv 邻域聚合基线」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphSAGE（src/onescience/models/cfd_benchmark/GraphSAGE.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、n_layers、space_dim。","1. 校验边无孤立异常。","2. 配置层数与通道。","3. forward 冒烟。","4. 训练。","5. 记录作为下限指标。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphsage-light-graph",title:"轻量图拓扑节点回归",summary:"在资源受限时用轻量图模型快速回归节点量",fields:[{...e[0],placeholder:"例如：已有网格邻接的非结构 CFD 节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_layers"},e[2],e[3]],capabilitySummary:{current:"轻量图拓扑节点回归",keyConfig:"batch、n_hidden、学习率",resources:"CPU/小 GPU",deliverables:"快速可跑结果、耗时"},requiredData:["图数据"],expectedOutputs:"快速基线包",promptTemplate:["请基于 {{system}} 体系完成 GraphSAGE「轻量图拓扑节点回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphSAGE（src/onescience/models/cfd_benchmark/GraphSAGE.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、n_layers、space_dim。","1. 缩小 n_hidden 冒烟。","2. 短训。","3. 评估。","4. 对比 PointNet 无图下限。","5. 决定是否升级 MeshGraphNet。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphsage-edge-pred",title:"edge_index 构图场预测",summary:"基于给定边索引做场预测并检查构图质量",fields:[{...e[0],placeholder:"例如：已有网格邻接的非结构 CFD 节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_layers"},e[2],e[3]],capabilitySummary:{current:"edge_index 构图场预测",keyConfig:"构图半径/网格边、space_dim",resources:"GPU",deliverables:"预测、坏边诊断"},requiredData:["可复现构图脚本","特征"],expectedOutputs:"预测与构图报告",promptTemplate:["请基于 {{system}} 体系完成 GraphSAGE「edge_index 构图场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphSAGE（src/onescience/models/cfd_benchmark/GraphSAGE.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、n_layers、space_dim。","1. 检查边对称与自环策略。","2. 训练/推理。","3. 高误差区对照连通性。","4. 必要时重构边。","5. 固化构图参数。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphsage-sage-floor",title:"GraphSAGE CFD 下限对比",summary:"作为图模型下限与 PointNet/MeshGraphNet 对比",fields:[{...e[0],placeholder:"例如：已有网格邻接的非结构 CFD 节点场"},{...e[1],placeholder:"例如：edge_index、节点特征、fun_dim/out_dim、n_layers"},e[2],e[3]],capabilitySummary:{current:"GraphSAGE CFD 下限对比",keyConfig:"同数据同划分、指标对齐",resources:"同协议对比",deliverables:"下限对比表"},requiredData:["统一 datapipe"],expectedOutputs:"对比结论",promptTemplate:["请基于 {{system}} 体系完成 GraphSAGE「GraphSAGE CFD 下限对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphSAGE（src/onescience/models/cfd_benchmark/GraphSAGE.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、n_layers、space_dim。","1. 锁定数据。","2. 跑 GraphSAGE。","3. 并列 PointNet/MGN 指标。","4. 解释邻域聚合收益。","5. 推荐下一主干。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},B={id:"graphViT",name:"GraphViT",category:"onemodel",domain:"fluid",tags:["流体","图Transformer","时序预报","簇池化"],blurb:"Eagle 网格图-视觉 Transformer：GNN+簇池化+全局注意力，自回归预报速度压力。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/graphvit",intro:{title:"30 秒看懂模型",subtitle:"在动态非结构网格上用簇级 latent graph 做长距离依赖，滚动预测节点速度与压力",suited:["Eagle 网格时序滚动预测","簇池化图 Transformer 训练","速度压力多步预报评估","cluster 文件与边界覆盖校验"],notFor:["纯规则图像网格","无时间维","无 cluster/节点类型且不改 datapipe"],prepare:["sim.npz/triangles","constrained_kmeans cluster","node_type one-hot","state_size"],obtain:["逐步状态预测","N-RMSE","动画可选","checkpoint"],steps:["核对 Eagle 协议","加载 cluster","配置 GraphViT","自回归训练","多步 N-RMSE 评估"]},install:{banner:"已为你创建 GraphViT 环境配置任务",prompt:["请为 GraphViT 创建一个可以直接跑通的环境配置任务。","模型定位：Eagle 网格 CFD 时序 GraphViT（GNN + RNNClusterPooling + Pre-LN Transformer）。","适合任务：Eagle 网格时序滚动预测、簇池化图 Transformer 训练、速度压力多步预报评估、cluster 文件与边界覆盖校验。","典型输入：sim.npz/triangles / constrained_kmeans cluster / node_type one-hot / state_size。","典型输出：逐步状态预测 / N-RMSE / 动画可选 / checkpoint。","关键配置：state_size、w_size、n_attention、nb_gn、n_heads、pos_length；数据侧 n_cluster、window_length_*、type_as_onehot。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/graphvit（配合 EagleDatapipe / conf/graphvit_eagle.yaml））与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：Eagle 网格时序滚动预测 / 簇池化图 Transformer 训练 / 速度压力多步预报评估。"].join(`
-`)},scenarios:[{id:"graphViT-eagle-roll",title:"Eagle 网格时序滚动预测",summary:"在 Eagle 网格上自回归滚动预测下一时刻节点状态",fields:[{...e[0],placeholder:"例如：Eagle 非结构网格二维流场时序"},{...e[1],placeholder:"例如：sim.npz、triangles.npy、constrained_kmeans_*.npy、graphvit_eagle.yaml"},e[2],e[3]],capabilitySummary:{current:"Eagle 网格时序滚动预测",keyConfig:"window_length、state_size、apply_noise、loss_alpha",resources:"远程 GPU；长窗口注意显存",deliverables:"滚动预测、逐步误差、checkpoint"},requiredData:["Eagle split 与 sim.npz","cluster 文件","node_type"],expectedOutputs:"时序预测与日志",promptTemplate:["请基于 {{system}} 体系完成 GraphViT「Eagle 网格时序滚动预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphViT（src/onescience/models/graphvit（配合 EagleDatapipe / conf/graphvit_eagle.yaml））执行，关键参数关注：state_size、w_size、n_attention、nb_gn、n_heads、pos_length；数据侧 n_cluster、window_length_*、type_as_onehot。","1. 检查 split/sim.npz/triangles/cluster。","2. 设 state_size=速度+压力维。","3. 配置 window 与 GraphViT 超参。","4. train_graphvit 短冒烟。","5. 正式训练并保存 best_model。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphViT-cluster-vit",title:"簇池化图 Transformer 训练",summary:"用 constrained k-means 簇池化后在簇图上做全局注意力训练",fields:[{...e[0],placeholder:"例如：Eagle 非结构网格二维流场时序"},{...e[1],placeholder:"例如：sim.npz、triangles.npy、constrained_kmeans_*.npy、graphvit_eagle.yaml"},e[2],e[3]],capabilitySummary:{current:"簇池化图 Transformer 训练",keyConfig:"n_cluster、w_size、n_attention、n_heads、nb_gn",resources:"GPU；n_cluster 与显存权衡",deliverables:"训练曲线、簇配置快照、权重"},requiredData:["匹配 n_cluster 的 cluster 文件","yaml"],expectedOutputs:"可复现训练配置",promptTemplate:["请基于 {{system}} 体系完成 GraphViT「簇池化图 Transformer 训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphViT（src/onescience/models/graphvit（配合 EagleDatapipe / conf/graphvit_eagle.yaml））执行，关键参数关注：state_size、w_size、n_attention、nb_gn、n_heads、pos_length；数据侧 n_cluster、window_length_*、type_as_onehot。","1. 确认 cluster 与 n_cluster 一致。","2. 检查 w_size+4*pos_length 可被 n_heads 整除。","3. 对齐 n_attention 与 mask。","4. 训练。","5. OOM 时增簇数而非硬拉长序列。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphViT-uvp-eval",title:"速度压力多步预报评估",summary:"加载最佳权重计算逐时间步速度/压力 N-RMSE",fields:[{...e[0],placeholder:"例如：Eagle 非结构网格二维流场时序"},{...e[1],placeholder:"例如：sim.npz、triangles.npy、constrained_kmeans_*.npy、graphvit_eagle.yaml"},e[2],e[3]],capabilitySummary:{current:"速度压力多步预报评估",keyConfig:"eval 窗口、denormalize、N-RMSE 步数、动画 cells",resources:"单卡评估",deliverables:"N-RMSE 曲线、可选流场动画、结论"},requiredData:["best_model.pth","测试 split","归一化统计"],expectedOutputs:"评估报告与可视化",promptTemplate:["请基于 {{system}} 体系完成 GraphViT「速度压力多步预报评估」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphViT（src/onescience/models/graphvit（配合 EagleDatapipe / conf/graphvit_eagle.yaml））执行，关键参数关注：state_size、w_size、n_attention、nb_gn、n_heads、pos_length；数据侧 n_cluster、window_length_*、type_as_onehot。","1. 加载 ckpt（兼容 DDP 前缀）。","2. apply_noise=False 评估。","3. 反归一化后算速度/压力指标。","4. 可选 with_cells 出动画。","5. 长程误差大则建议加窗或 scheduled sampling。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphViT-cluster-bc",title:"cluster 文件与边界覆盖校验",summary:"校验簇分配、节点类型与墙面/入口边界状态覆盖",fields:[{...e[0],placeholder:"例如：Eagle 非结构网格二维流场时序"},{...e[1],placeholder:"例如：sim.npz、triangles.npy、constrained_kmeans_*.npy、graphvit_eagle.yaml"},e[2],e[3]],capabilitySummary:{current:"cluster 文件与边界覆盖校验",keyConfig:"node_type 语义、clusters_mask、边界覆盖逻辑",resources:"登录节点即可诊断",deliverables:"字段 shape 报告、边界覆盖检查表"},requiredData:["clusters/clusters_mask","node_type","样例 batch"],expectedOutputs:"协议诊断与修复清单",promptTemplate:["请基于 {{system}} 体系完成 GraphViT「cluster 文件与边界覆盖校验」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphViT（src/onescience/models/graphvit（配合 EagleDatapipe / conf/graphvit_eagle.yaml））执行，关键参数关注：state_size、w_size、n_attention、nb_gn、n_heads、pos_length；数据侧 n_cluster、window_length_*、type_as_onehot。","1. 打印 mesh_pos/edges/state/node_type/clusters shape。","2. 确认 node_type 9 维 one-hot。","3. 检查墙/入口/禁用节点覆盖。","4. 缺 cluster 则先 clusterize。","5. 通过后再启动训练。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},j={id:"lsm",name:"LSM",category:"onemodel",domain:"fluid",tags:["流体","潜在谱模型","混合算子","多尺度"],blurb:"潜在谱模型：谱卷积+Transformer+U 型结构在隐空间学习 CFD 算子的强基线。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/LSM.py",intro:{title:"30 秒看懂模型",subtitle:"希望同时比较谱、注意力与多尺度局部结构时，用 LSM 作高容量混合算子",suited:["潜在谱混合算子训练","谱+注意力+U型联合建模","LSM 强基线场预测","多组件超参排错诊断"],notFor:["资源很紧只要轻量基线","排错预算不足","数据协议未稳"],prepare:["规则/投影网格","modes、n_heads","fun_dim/out_dim/task","shapelist"],obtain:["强基线预测","组件消融","checkpoint"],steps:["补齐 task/modes","混合结构冒烟","训练","组件消融","排错清单"]},install:{banner:"已为你创建 LSM 环境配置任务",prompt:["请为 LSM 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Latent Spectral Model。","适合任务：潜在谱混合算子训练、谱+注意力+U型联合建模、LSM 强基线场预测、多组件超参排错诊断。","典型输入：规则/投影网格 / modes、n_heads / fun_dim/out_dim/task / shapelist。","典型输出：强基线预测 / 组件消融 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、modes、n_heads、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/LSM.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：潜在谱混合算子训练 / 谱+注意力+U型联合建模 / LSM 强基线场预测。"].join(`
-`)},scenarios:[{id:"lsm-latent-mix",title:"潜在谱混合算子训练",summary:"在潜在空间联合谱卷积与注意力训练混合算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"潜在谱混合算子训练",keyConfig:"modes、n_heads、n_hidden、task、fun_dim、out_dim",resources:"远程大显存 GPU",deliverables:"checkpoint、曲线、相对 L2"},requiredData:["完整 LSM args","稳定 datapipe"],expectedOutputs:"强基线权重",promptTemplate:["请基于 {{system}} 体系完成 LSM「潜在谱混合算子训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience LSM（src/onescience/models/cfd_benchmark/LSM.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_heads、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 补齐含 task 的全部字段。","2. 假 batch 冒烟。","3. 短训确认各组件可反传。","4. 正式训练。","5. 保存最佳。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"lsm-spec-attn-u",title:"谱+注意力+U型联合建模",summary:"启用谱、注意力与 U 型多尺度路径的联合建模",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"谱+注意力+U型联合建模",keyConfig:"各分支开关/宽度、shapelist、modes",resources:"高显存",deliverables:"联合模型、分支贡献初探"},requiredData:["结构网格","标签"],expectedOutputs:"联合训练结果",promptTemplate:["请基于 {{system}} 体系完成 LSM「谱+注意力+U型联合建模」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience LSM（src/onescience/models/cfd_benchmark/LSM.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_heads、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 确认 shapelist。","2. 配置三组件容量。","3. 训练。","4. 监控各 loss 项。","5. OOM 先降 n_hidden/modes。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"lsm-lsm-strong",title:"LSM 强基线场预测",summary:"作为强基线在测试集上给出场预测上限参考",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"LSM 强基线场预测",keyConfig:"checkpoint、反归一化、相对 L2",resources:"单卡推理",deliverables:"预测场、基线分数"},requiredData:["训练好的 LSM","测试集"],expectedOutputs:"基线分数卡",promptTemplate:["请基于 {{system}} 体系完成 LSM「LSM 强基线场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience LSM（src/onescience/models/cfd_benchmark/LSM.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_heads、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 加载权重。","2. 推理评估。","3. 与 FNO/U_NO 对比。","4. 记录优势场景。","5. 作为后续模型对照。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"lsm-lsm-debug",title:"多组件超参排错诊断",summary:"当训练不稳时按组件隔离定位 modes/注意力/U 路径问题",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"多组件超参排错诊断",keyConfig:"逐组件关闭、学习率、归一化",resources:"短训诊断",deliverables:"故障树、最小可跑配置"},requiredData:["复现失败日志","可改 YAML"],expectedOutputs:"排错报告",promptTemplate:["请基于 {{system}} 体系完成 LSM「多组件超参排错诊断」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience LSM（src/onescience/models/cfd_benchmark/LSM.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_heads、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 复现失败。","2. 逐一简化组件。","3. 定位坏配置。","4. 给出最小稳定超参。","5. 再逐步加回容量。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},L={id:"meshgraphnet",name:"MeshGraphNet",category:"onemodel",domain:"fluid",tags:["流体","网格图","消息传递","MeshGraphNet"],blurb:"Encode-Process-Decode 网格图网络：显式边特征消息传递，预测非结构节点物理量。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/cfd/MeshGraphNet",intro:{title:"30 秒看懂模型",subtitle:"已有 DGL/CuGraph 与边特征时，用多层 Edge→Node 消息传递做网格节点级代理",suited:["Encode-Process-Decode 网格训练","边特征消息传递场预测","MeshGraphNet 非结构推理","DGL/CuGraph 图协议对齐"],notFor:["只有 (x,fx) 无 graph/edge_features","模型不负责构图","无网格拓扑"],prepare:["DGLGraph/CuGraphCSC","edge_features","fun_dim/out_dim","节点目标"],obtain:["节点场","消息传递诊断","checkpoint"],steps:["对齐图协议","设 fun_dim/out_dim","processor 冒烟","训练","非结构推理"]},install:{banner:"已为你创建 MeshGraphNet 环境配置任务",prompt:["请为 MeshGraphNet 创建一个可以直接跑通的环境配置任务。","模型定位：MeshGraphNet encode-process-decode 图消息传递（examples/cfd/MeshGraphNet）。","适合任务：Encode-Process-Decode 网格训练、边特征消息传递场预测、MeshGraphNet 非结构推理、DGL/CuGraph 图协议对齐。","典型输入：DGLGraph/CuGraphCSC / edge_features / fun_dim/out_dim / 节点目标。","典型输出：节点场 / 消息传递诊断 / checkpoint。","关键配置：fun_dim、out_dim；图侧 node/edge 特征与 processor 步数（示例配置）。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（examples/cfd/MeshGraphNet）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：Encode-Process-Decode 网格训练 / 边特征消息传递场预测 / MeshGraphNet 非结构推理。"].join(`
-`)},scenarios:[{id:"meshgraphnet-epd-mesh",title:"Encode-Process-Decode 网格训练",summary:"按 EPD 范式训练网格节点场代理",fields:[{...e[0],placeholder:"例如：圆柱绕流/翼型非结构网格瞬时场"},{...e[1],placeholder:"例如：MeshGraphNet 示例数据、DGL 图、edge_features、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"Encode-Process-Decode 网格训练",keyConfig:"fun_dim、out_dim、processor 步数、学习率",resources:"远程 GPU；大图注意分 batch",deliverables:"checkpoint、曲线、节点误差"},requiredData:["图+边特征+标签","示例入口"],expectedOutputs:"训练权重",promptTemplate:["请基于 {{system}} 体系完成 MeshGraphNet「Encode-Process-Decode 网格训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MeshGraphNet（examples/cfd/MeshGraphNet）执行，关键参数关注：fun_dim、out_dim；图侧 node/edge 特征与 processor 步数（示例配置）。","1. 核对 examples/cfd/MeshGraphNet 数据。","2. 对齐 fun_dim/out_dim。","3. encode-process-decode 冒烟。","4. 训练。","5. 保存 ckpt。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"meshgraphnet-edge-mp",title:"边特征消息传递场预测",summary:"利用显式边特征做 Edge→Node 消息传递预测",fields:[{...e[0],placeholder:"例如：圆柱绕流/翼型非结构网格瞬时场"},{...e[1],placeholder:"例如：MeshGraphNet 示例数据、DGL 图、edge_features、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"边特征消息传递场预测",keyConfig:"edge_features 维、消息步数、归一化",resources:"GPU",deliverables:"预测场、边特征消融"},requiredData:["含边特征的图"],expectedOutputs:"预测与消融",promptTemplate:["请基于 {{system}} 体系完成 MeshGraphNet「边特征消息传递场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MeshGraphNet（examples/cfd/MeshGraphNet）执行，关键参数关注：fun_dim、out_dim；图侧 node/edge 特征与 processor 步数（示例配置）。","1. 检查边特征维。","2. 训练/推理。","3. 边特征置零对照。","4. 解释边信息贡献。","5. 固化特征定义。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"meshgraphnet-mgn-infer",title:"MeshGraphNet 非结构推理",summary:"加载权重在新网格上做节点场推理",fields:[{...e[0],placeholder:"例如：圆柱绕流/翼型非结构网格瞬时场"},{...e[1],placeholder:"例如：MeshGraphNet 示例数据、DGL 图、edge_features、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"MeshGraphNet 非结构推理",keyConfig:"checkpoint、图构造一致、反归一化",resources:"单卡",deliverables:"节点预测、误差、失败图列表"},requiredData:["权重","同协议构图"],expectedOutputs:"推理结果",promptTemplate:["请基于 {{system}} 体系完成 MeshGraphNet「MeshGraphNet 非结构推理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MeshGraphNet（examples/cfd/MeshGraphNet）执行，关键参数关注：fun_dim、out_dim；图侧 node/edge 特征与 processor 步数（示例配置）。","1. 同训练协议构图。","2. 加载权重。","3. 前向。","4. 评估。","5. 协议不一致则先修 datapipe。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"meshgraphnet-dgl-align",title:"DGL/CuGraph 图协议对齐",summary:"对齐 DGLGraph/CuGraphCSC 字段与模型输入约定",fields:[{...e[0],placeholder:"例如：圆柱绕流/翼型非结构网格瞬时场"},{...e[1],placeholder:"例如：MeshGraphNet 示例数据、DGL 图、edge_features、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"DGL/CuGraph 图协议对齐",keyConfig:"图后端、CSR/CSC、特征字段名",resources:"登录节点诊断+GPU 冒烟",deliverables:"协议对照表、最小可跑 batch"},requiredData:["样例图","模型 forward 签名"],expectedOutputs:"对齐清单",promptTemplate:["请基于 {{system}} 体系完成 MeshGraphNet「DGL/CuGraph 图协议对齐」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MeshGraphNet（examples/cfd/MeshGraphNet）执行，关键参数关注：fun_dim、out_dim；图侧 node/edge 特征与 processor 步数（示例配置）。","1. 打印图字段与 shape。","2. 对照 Model 期望输入。","3. 修正转换层。","4. 冒烟。","5. 写入数据卡。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},R={id:"mwt",name:"MWT",category:"onemodel",domain:"fluid",tags:["流体","多小波","算子学习","规则网格"],blurb:"多小波变换算子：在多尺度频域建模，结构分支常把空间尺寸 pad 到 2 的幂。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/MWT.py",intro:{title:"30 秒看懂模型",subtitle:"规则网格需要多尺度频域表达且可接受 2 幂 padding 时的算子学习主干",suited:["多小波变换算子训练","2幂 padding 多尺度频域拟合","MWT 规则网格场代理","mwt_k/modes 小波截断调参"],notFor:["shapelist 不一致","padding 后边界敏感","非规则且无法 pad"],prepare:["规则网格","modes、mwt_k","fun_dim/out_dim","shapelist"],obtain:["多尺度预测","padding 说明","checkpoint"],steps:["检查 2 幂 pad","设 mwt_k/modes","冒烟","训练","边界误差检查"]},install:{banner:"已为你创建 MWT 环境配置任务",prompt:["请为 MWT 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Multiwavelet Transform operator。","适合任务：多小波变换算子训练、2幂 padding 多尺度频域拟合、MWT 规则网格场代理、mwt_k/modes 小波截断调参。","典型输入：规则网格 / modes、mwt_k / fun_dim/out_dim / shapelist。","典型输出：多尺度预测 / padding 说明 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、modes、mwt_k、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/MWT.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：多小波变换算子训练 / 2幂 padding 多尺度频域拟合 / MWT 规则网格场代理。"].join(`
-`)},scenarios:[{id:"mwt-mwt-train",title:"多小波变换算子训练",summary:"用多小波 Fourier 组件训练规则网格算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"多小波变换算子训练",keyConfig:"mwt_k、modes、n_layers、n_hidden、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["规则网格数据","完整 args"],expectedOutputs:"权重与指标",promptTemplate:["请基于 {{system}} 体系完成 MWT「多小波变换算子训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MWT（src/onescience/models/cfd_benchmark/MWT.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、mwt_k、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 确认 geotype/shapelist。","2. 配置 mwt_k/modes。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"mwt-pow2-pad",title:"2幂 padding 多尺度频域拟合",summary:"处理空间尺寸 pad 到 2 幂后的多尺度拟合与裁剪",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"2幂 padding 多尺度频域拟合",keyConfig:"pad 策略、裁剪回原尺寸、边界处理",resources:"GPU",deliverables:"pad/crop 校验报告、训练结果"},requiredData:["原始 H/W 与 pad 后尺寸说明"],expectedOutputs:"拟合结果与边界说明",promptTemplate:["请基于 {{system}} 体系完成 MWT「2幂 padding 多尺度频域拟合」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MWT（src/onescience/models/cfd_benchmark/MWT.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、mwt_k、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 记录 pad 前后尺寸。","2. 训练。","3. crop 回原网格评估。","4. 检查边界条带误差。","5. 敏感则改尺寸或换 FNO。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"mwt-mwt-surr",title:"MWT 规则网格场代理",summary:"部署 MWT 作为规则网格场代理做推理",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"MWT 规则网格场代理",keyConfig:"checkpoint、反归一化、modes",resources:"单卡",deliverables:"代理场、误差图"},requiredData:["权重","测试网格"],expectedOutputs:"代理结果",promptTemplate:["请基于 {{system}} 体系完成 MWT「MWT 规则网格场代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MWT（src/onescience/models/cfd_benchmark/MWT.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、mwt_k、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 加载权重。","2. 推理 crop。","3. 评估。","4. 多尺度误差分析。","5. 给出 modes/mwt_k 调整建议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"mwt-mwt-tune",title:"mwt_k/modes 小波截断调参",summary:"扫描小波阶与谱截断平衡精度算力",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"mwt_k/modes 小波截断调参",keyConfig:"mwt_k×modes 网格搜索",resources:"短训搜索",deliverables:"调参曲线、推荐对"},requiredData:["验证集"],expectedOutputs:"推荐超参",promptTemplate:["请基于 {{system}} 体系完成 MWT「mwt_k/modes 小波截断调参」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MWT（src/onescience/models/cfd_benchmark/MWT.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、mwt_k、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 设计小搜索空间。","2. 短训。","3. 画误差-耗时。","4. 选推荐对。","5. 固化到 YAML。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},q={id:"ono",name:"ONO",category:"onemodel",domain:"fluid",tags:["流体","正交算子","注意力近似","低秩"],blurb:"正交/低秩风格神经算子：可切换 nystrom/linear/selfAttention 做点级场预测。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/ONO.py",intro:{title:"30 秒看懂模型",subtitle:"需要比较多种算子注意力近似时，用坐标/场双分支 + ONO block 做点级回归",suited:["正交神经算子注意力训练","nystrom/linear 注意力对比","ONO 低秩算子场回归","psi_dim/attn_type 配置校验"],notFor:["缺双分支输入","psi_dim/attn_type 未配置","只要谱卷积"],prepare:["坐标与 fx","attn_type、psi_dim","n_heads/n_layers","fun_dim/out_dim"],obtain:["点级场","attn_type 对比表","checkpoint"],steps:["设 attn_type/psi_dim","双分支冒烟","训练","近似对比","配置校验"]},install:{banner:"已为你创建 ONO 环境配置任务",prompt:["请为 ONO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Orthogonal Neural Operator。","适合任务：正交神经算子注意力训练、nystrom/linear 注意力对比、ONO 低秩算子场回归、psi_dim/attn_type 配置校验。","典型输入：坐标与 fx / attn_type、psi_dim / n_heads/n_layers / fun_dim/out_dim。","典型输出：点级场 / attn_type 对比表 / checkpoint。","关键配置：act、attn_type、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、psi_dim、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/ONO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：正交神经算子注意力训练 / nystrom/linear 注意力对比 / ONO 低秩算子场回归。"].join(`
-`)},scenarios:[{id:"ono-ono-attn",title:"正交神经算子注意力训练",summary:"配置 ONO block 与注意力类型训练点级算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"正交神经算子注意力训练",keyConfig:"attn_type、psi_dim、n_heads、n_layers、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["双分支输入","显式 attn_type/psi_dim"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 ONO「正交神经算子注意力训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience ONO（src/onescience/models/cfd_benchmark/ONO.py）执行，关键参数关注：act、attn_type、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、psi_dim、ref、shapelist、space_dim、time_input、unified_pos。","1. 写入 attn_type 与 psi_dim。","2. 补齐其余 args。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"ono-nystrom-linear",title:"nystrom/linear 注意力对比",summary:"在同数据上对比不同 attn_type 的精度与耗时",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"nystrom/linear 注意力对比",keyConfig:"attn_type∈{nystrom,linear,selfAttention}、同种子",resources:"同卡对比",deliverables:"对比表、推荐类型"},requiredData:["可切换配置"],expectedOutputs:"对比报告",promptTemplate:["请基于 {{system}} 体系完成 ONO「nystrom/linear 注意力对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience ONO（src/onescience/models/cfd_benchmark/ONO.py）执行，关键参数关注：act、attn_type、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、psi_dim、ref、shapelist、space_dim、time_input、unified_pos。","1. 固定数据与层数。","2. 切换 attn_type 训练/评估。","3. 记录误差与步时。","4. 推荐默认类型。","5. 写入配置说明。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"ono-ono-lowrank",title:"ONO 低秩算子场回归",summary:"利用正交/低秩近似做点级物理场回归",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"ONO 低秩算子场回归",keyConfig:"psi_dim、n_hidden、dropout",resources:"GPU",deliverables:"回归场、低秩维敏感性"},requiredData:["点级数据"],expectedOutputs:"回归结果",promptTemplate:["请基于 {{system}} 体系完成 ONO「ONO 低秩算子场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience ONO（src/onescience/models/cfd_benchmark/ONO.py）执行，关键参数关注：act、attn_type、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、psi_dim、ref、shapelist、space_dim、time_input、unified_pos。","1. 扫描 psi_dim。","2. 训练。","3. 评估。","4. 选稳定维。","5. 固化。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"ono-psi-attn-cfg",title:"psi_dim/attn_type 配置校验",summary:"在训练前校验关键注意力配置完整性",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"psi_dim/attn_type 配置校验",keyConfig:"必填字段清单、假输入",resources:"本地冒烟",deliverables:"配置检查表、可实例化 args"},requiredData:["草稿 YAML"],expectedOutputs:"校验通过的配置",promptTemplate:["请基于 {{system}} 体系完成 ONO「psi_dim/attn_type 配置校验」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience ONO（src/onescience/models/cfd_benchmark/ONO.py）执行，关键参数关注：act、attn_type、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、psi_dim、ref、shapelist、space_dim、time_input、unified_pos。","1. 列出源码读取字段。","2. 补齐缺失。","3. 实例化 Model。","4. 假 batch。","5. 通过后再提交训练。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},I={id:"pointnet",name:"PointNet",category:"onemodel",domain:"fluid",tags:["流体","点云","MLP基线","非结构"],blurb:"点云共享 MLP 基线：不显式用邻接，快速建立非结构点集场回归下限。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/PointNet.py",intro:{title:"30 秒看懂模型",subtitle:"图拓扑暂不可用或需要快速下限时，用坐标+特征共享 MLP 做点级回归",suited:["点云 MLP 下限建模","无拓扑点级场回归","PointNet 快速冒烟基线","非结构点集通道对齐"],notFor:["强依赖局部邻域","需要边界连通性","长程相互作用主导"],prepare:["点坐标","点特征 fun_dim","out_dim、space_dim","n_hidden"],obtain:["点级预测","下限指标","checkpoint"],steps:["对齐通道","设 n_hidden","冒烟","短训下限","对比图模型"]},install:{banner:"已为你创建 PointNet 环境配置任务",prompt:["请为 PointNet 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark PointNet 点云 MLP 基线。","适合任务：点云 MLP 下限建模、无拓扑点级场回归、PointNet 快速冒烟基线、非结构点集通道对齐。","典型输入：点坐标 / 点特征 fun_dim / out_dim、space_dim / n_hidden。","典型输出：点级预测 / 下限指标 / checkpoint。","关键配置：act、fun_dim、out_dim、n_hidden、space_dim。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/PointNet.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：点云 MLP 下限建模 / 无拓扑点级场回归 / PointNet 快速冒烟基线。"].join(`
-`)},scenarios:[{id:"pointnet-pc-mlp",title:"点云 MLP 下限建模",summary:"用共享 MLP 训练点云场回归下限模型",fields:[{...e[0],placeholder:"例如：无边索引的非结构点云 CFD 场"},{...e[1],placeholder:"例如：点坐标、点特征、fun_dim/out_dim、space_dim"},e[2],e[3]],capabilitySummary:{current:"点云 MLP 下限建模",keyConfig:"n_hidden、fun_dim、out_dim、act、space_dim",resources:"单卡/CPU 均可短训",deliverables:"checkpoint、下限相对误差"},requiredData:["点坐标+特征+标签"],expectedOutputs:"下限权重",promptTemplate:["请基于 {{system}} 体系完成 PointNet「点云 MLP 下限建模」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience PointNet（src/onescience/models/cfd_benchmark/PointNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、space_dim。","1. 拼接坐标与特征。","2. 配置 MLP 宽度。","3. 冒烟。","4. 短训。","5. 记录下限分。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"pointnet-notopo-reg",title:"无拓扑点级场回归",summary:"在无 edge_index 场景完成点级场回归",fields:[{...e[0],placeholder:"例如：无边索引的非结构点云 CFD 场"},{...e[1],placeholder:"例如：点坐标、点特征、fun_dim/out_dim、space_dim"},e[2],e[3]],capabilitySummary:{current:"无拓扑点级场回归",keyConfig:"归一化、out_dim、batch 点数",resources:"GPU 可选",deliverables:"点级预测、误差"},requiredData:["点集数据"],expectedOutputs:"回归结果",promptTemplate:["请基于 {{system}} 体系完成 PointNet「无拓扑点级场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience PointNet（src/onescience/models/cfd_benchmark/PointNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、space_dim。","1. 确认无图依赖。","2. 训练/推理。","3. 评估。","4. 误差大则建议 RegDGCNN/图模型。","5. 写入选型说明。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"pointnet-pn-smoke",title:"PointNet 快速冒烟基线",summary:"最小配置验证 datapipe 与 out_dim 对接",fields:[{...e[0],placeholder:"例如：无边索引的非结构点云 CFD 场"},{...e[1],placeholder:"例如：点坐标、点特征、fun_dim/out_dim、space_dim"},e[2],e[3]],capabilitySummary:{current:"PointNet 快速冒烟基线",keyConfig:"假点云、fun_dim/out_dim",resources:"本地冒烟",deliverables:"shape 报告、可跑证明"},requiredData:["args 草稿"],expectedOutputs:"冒烟报告",promptTemplate:["请基于 {{system}} 体系完成 PointNet「PointNet 快速冒烟基线」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience PointNet（src/onescience/models/cfd_benchmark/PointNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、space_dim。","1. 构造假点云。","2. 实例化。","3. 检查输出维。","4. 接入真实一小批。","5. 再扩训练。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"pointnet-pc-align",title:"非结构点集通道对齐",summary:"对齐 space_dim/fun_dim/out_dim 与点特征布局",fields:[{...e[0],placeholder:"例如：无边索引的非结构点云 CFD 场"},{...e[1],placeholder:"例如：点坐标、点特征、fun_dim/out_dim、space_dim"},e[2],e[3]],capabilitySummary:{current:"非结构点集通道对齐",keyConfig:"通道布局、坐标维、特征维",resources:"诊断为主",deliverables:"通道数据卡、修正后的 args"},requiredData:["原始点特征说明"],expectedOutputs:"对齐后的配置",promptTemplate:["请基于 {{system}} 体系完成 PointNet「非结构点集通道对齐」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience PointNet（src/onescience/models/cfd_benchmark/PointNet.py）执行，关键参数关注：act、fun_dim、out_dim、n_hidden、space_dim。","1. 打印输入最后一维。","2. 对照源码期望。","3. 修正 fun_dim/space_dim。","4. 冒烟。","5. 固化数据卡。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},H={id:"regdgcnn",name:"RegDGCNN",category:"onemodel",domain:"fluid",tags:["流体","动态图","EdgeConv","点云"],blurb:"动态图卷积点云回归：按特征动态构图做 EdgeConv，捕捉局部几何邻域。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/RegDGCNN.py",intro:{title:"30 秒看懂模型",subtitle:"无固定边但局部邻域重要时，用动态 kNN/EdgeConv 做点云节点级场回归",suited:["动态图 EdgeConv 回归","局部邻域点云场预测","RegDGCNN 动态构图训练","kNN 邻域敏感性评估"],notFor:["期望使用外部 geo 图（源码忽略 geo）","只需全局谱","完全无局部几何"],prepare:["点坐标/特征","emb_dims、dropout","fun_dim/out_dim/space_dim","n_hidden"],obtain:["动态图预测","k 敏感性","checkpoint"],steps:["确认动态构图","设 emb_dims","冒烟","训练","扫邻域 k"]},install:{banner:"已为你创建 RegDGCNN 环境配置任务",prompt:["请为 RegDGCNN 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark RegDGCNN 动态图点云回归。","适合任务：动态图 EdgeConv 回归、局部邻域点云场预测、RegDGCNN 动态构图训练、kNN 邻域敏感性评估。","典型输入：点坐标/特征 / emb_dims、dropout / fun_dim/out_dim/space_dim / n_hidden。","典型输出：动态图预测 / k 敏感性 / checkpoint。","关键配置：dropout、emb_dims、fun_dim、out_dim、n_hidden、space_dim。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/RegDGCNN.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：动态图 EdgeConv 回归 / 局部邻域点云场预测 / RegDGCNN 动态构图训练。"].join(`
-`)},scenarios:[{id:"regdgcnn-edgeconv-reg",title:"动态图 EdgeConv 回归",summary:"基于动态构图与 EdgeConv 训练点云回归",fields:[{...e[0],placeholder:"例如：汽车表面/翼型点云压力场"},{...e[1],placeholder:"例如：点云坐标特征、emb_dims、fun_dim/out_dim、dropout"},e[2],e[3]],capabilitySummary:{current:"动态图 EdgeConv 回归",keyConfig:"emb_dims、n_hidden、dropout、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、相对误差"},requiredData:["点云+标签"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 RegDGCNN「动态图 EdgeConv 回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience RegDGCNN（src/onescience/models/cfd_benchmark/RegDGCNN.py）执行，关键参数关注：dropout、emb_dims、fun_dim、out_dim、n_hidden、space_dim。","1. 配置 emb_dims 等。","2. 冒烟（忽略外部 geo）。","3. 训练。","4. 评估。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"regdgcnn-local-pc",title:"局部邻域点云场预测",summary:"强调局部几何邻域的点云物理场预测",fields:[{...e[0],placeholder:"例如：汽车表面/翼型点云压力场"},{...e[1],placeholder:"例如：点云坐标特征、emb_dims、fun_dim/out_dim、dropout"},e[2],e[3]],capabilitySummary:{current:"局部邻域点云场预测",keyConfig:"邻域规模、归一化、space_dim",resources:"GPU",deliverables:"预测场、局部误差热区"},requiredData:["点云测试集"],expectedOutputs:"预测与热区",promptTemplate:["请基于 {{system}} 体系完成 RegDGCNN「局部邻域点云场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience RegDGCNN（src/onescience/models/cfd_benchmark/RegDGCNN.py）执行，关键参数关注：dropout、emb_dims、fun_dim、out_dim、n_hidden、space_dim。","1. 推理。","2. 可视化局部误差。","3. 对照曲率/边界。","4. 调整容量。","5. 结论是否够用。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"regdgcnn-dyn-graph-train",title:"RegDGCNN 动态构图训练",summary:"完整训练流程并记录动态图行为",fields:[{...e[0],placeholder:"例如：汽车表面/翼型点云压力场"},{...e[1],placeholder:"例如：点云坐标特征、emb_dims、fun_dim/out_dim、dropout"},e[2],e[3]],capabilitySummary:{current:"RegDGCNN 动态构图训练",keyConfig:"学习率、batch、dropout",resources:"GPU",deliverables:"曲线、构图耗时、权重"},requiredData:["训练配置"],expectedOutputs:"训练包",promptTemplate:["请基于 {{system}} 体系完成 RegDGCNN「RegDGCNN 动态构图训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience RegDGCNN（src/onescience/models/cfd_benchmark/RegDGCNN.py）执行，关键参数关注：dropout、emb_dims、fun_dim、out_dim、n_hidden、space_dim。","1. 启动训练。","2. 记录构图耗时。","3. 监控过拟合。","4. 早停。","5. 导出。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"regdgcnn-knn-sens",title:"kNN 邻域敏感性评估",summary:"评估动态邻域规模对精度与耗时的影响",fields:[{...e[0],placeholder:"例如：汽车表面/翼型点云压力场"},{...e[1],placeholder:"例如：点云坐标特征、emb_dims、fun_dim/out_dim、dropout"},e[2],e[3]],capabilitySummary:{current:"kNN 邻域敏感性评估",keyConfig:"邻域 k 或等价超参扫描",resources:"短训扫描",deliverables:"k-误差/耗时曲线"},requiredData:["可改邻域配置"],expectedOutputs:"推荐邻域设置",promptTemplate:["请基于 {{system}} 体系完成 RegDGCNN「kNN 邻域敏感性评估」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience RegDGCNN（src/onescience/models/cfd_benchmark/RegDGCNN.py）执行，关键参数关注：dropout、emb_dims、fun_dim、out_dim、n_hidden、space_dim。","1. 扫描邻域规模。","2. 短评估。","3. 选推荐 k。","4. 写入配置。","5. 提醒与固定图模型差异。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},z={id:"swin_transformer",name:"Swin Transformer",category:"onemodel",domain:"fluid",tags:["流体","窗口注意力","structured_2D","Swin"],blurb:"窗口注意力 Transformer：仅支持 structured_2D，适合规则二维网格局部-全局建模。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/Swin_Transformer.py",intro:{title:"30 秒看懂模型",subtitle:"规则二维网格上比较窗口局部注意力与全局注意力/谱算子时的场建模选择",suited:["窗口注意力 2D 场建模","structured_2D Swin 训练","局部窗口与全局谱对比","窗口 padding 网格对齐"],notFor:["非 structured_2D","1D/3D/非结构（会抛错）","无法 pad 到窗口倍数"],prepare:["structured_2D 场","shapelist","n_heads/n_layers","fun_dim/out_dim"],obtain:["2D 预测场","窗口对齐说明","checkpoint"],steps:["确认 structured_2D","pad 到窗口倍","冒烟","训练","与 FNO 对比"]},install:{banner:"已为你创建 Swin Transformer 环境配置任务",prompt:["请为 Swin Transformer 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark Swin Transformer（仅 structured_2D）。","适合任务：窗口注意力 2D 场建模、structured_2D Swin 训练、局部窗口与全局谱对比、窗口 padding 网格对齐。","典型输入：structured_2D 场 / shapelist / n_heads/n_layers / fun_dim/out_dim。","典型输出：2D 预测场 / 窗口对齐说明 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/Swin_Transformer.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：窗口注意力 2D 场建模 / structured_2D Swin 训练 / 局部窗口与全局谱对比。"].join(`
-`)},scenarios:[{id:"swin_transformer-win-attn-2d",title:"窗口注意力 2D 场建模",summary:"在 2D 规则网格上用窗口注意力建模物理场",fields:[{...e[0],placeholder:"例如：规则二维 CFD 场、窗内局部结构明显"},{...e[1],placeholder:"例如：structured_2D、shapelist、n_heads、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"窗口注意力 2D 场建模",keyConfig:"n_heads、n_layers、n_hidden、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["geotype=structured_2D","shapelist"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 Swin Transformer「窗口注意力 2D 场建模」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Swin Transformer（src/onescience/models/cfd_benchmark/Swin_Transformer.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 断言 structured_2D。","2. 配置窗口注意力超参。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"swin_transformer-swin-train",title:"structured_2D Swin 训练",summary:"完整 Swin 训练流程含 padding 与裁剪",fields:[{...e[0],placeholder:"例如：规则二维 CFD 场、窗内局部结构明显"},{...e[1],placeholder:"例如：structured_2D、shapelist、n_heads、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"structured_2D Swin 训练",keyConfig:"window pad、shapelist、batch",resources:"GPU",deliverables:"训练日志、pad 说明、权重"},requiredData:["2D 数据"],expectedOutputs:"可复现训练",promptTemplate:["请基于 {{system}} 体系完成 Swin Transformer「structured_2D Swin 训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Swin Transformer（src/onescience/models/cfd_benchmark/Swin_Transformer.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 计算 pad 到窗口倍数。","2. 训练。","3. crop 评估。","4. 记录边界影响。","5. 固化预处理。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"swin_transformer-win-vs-spec",title:"局部窗口与全局谱对比",summary:"对比 Swin 窗口注意力与 FNO 谱卷积",fields:[{...e[0],placeholder:"例如：规则二维 CFD 场、窗内局部结构明显"},{...e[1],placeholder:"例如：structured_2D、shapelist、n_heads、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"局部窗口与全局谱对比",keyConfig:"同数据、精度/显存/局部误差",resources:"同卡对比",deliverables:"对比表"},requiredData:["FNO 与 Swin 配置"],expectedOutputs:"选型建议",promptTemplate:["请基于 {{system}} 体系完成 Swin Transformer「局部窗口与全局谱对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Swin Transformer（src/onescience/models/cfd_benchmark/Swin_Transformer.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 固定划分。","2. 两模型评估。","3. 比较局部细节误差。","4. 给出场景推荐。","5. 记录算力差。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"swin_transformer-win-pad",title:"窗口 padding 网格对齐",summary:"校验 H/W 与窗口大小对齐及 pad/crop 正确性",fields:[{...e[0],placeholder:"例如：规则二维 CFD 场、窗内局部结构明显"},{...e[1],placeholder:"例如：structured_2D、shapelist、n_heads、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"窗口 padding 网格对齐",keyConfig:"窗口大小、pad/crop、shapelist",resources:"本地诊断",deliverables:"对齐检查表、示例 tensor shape"},requiredData:["原始分辨率","窗口配置"],expectedOutputs:"对齐通过证明",promptTemplate:["请基于 {{system}} 体系完成 Swin Transformer「窗口 padding 网格对齐」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Swin Transformer（src/onescience/models/cfd_benchmark/Swin_Transformer.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 打印 pad 前后 shape。","2. 跑一层 block。","3. 确认 crop 回原尺寸。","4. 写入约束。","5. 再训练。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},W={id:"transformer",name:"Transformer",category:"onemodel",domain:"fluid",tags:["流体","自注意力","Transformer","算子基线"],blurb:"标准自注意力 Transformer 算子基线；EagleMesh 示例可用于网格场全局交互。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/cfd/EagleMeshTransformer",intro:{title:"30 秒看懂模型",subtitle:"点数适中时用普通全局自注意力作与 CFD_Benchmark 接口一致的场回归基线",suited:["全局自注意力算子基线","Eagle/Mesh Transformer 场回归","标准 Transformer CFD 对比","适中点数全局交互训练"],notFor:["NumPoints 很大无稀疏化","显存随 token 平方涨不可接受","更适合窗口/因子化注意力"],prepare:["token 化 (x,fx)","n_heads/n_layers","fun_dim/out_dim","点数预算"],obtain:["全局注意力预测","显存曲线","checkpoint"],steps:["估点数","设注意力超参","冒烟","训练","与 Galerkin/Factformer 对比"]},install:{banner:"已为你创建 Transformer 环境配置任务",prompt:["请为 Transformer 创建一个可以直接跑通的环境配置任务。","模型定位：标准自注意力 Transformer（CFD_Benchmark / examples/cfd/EagleMeshTransformer）。","适合任务：全局自注意力算子基线、Eagle/Mesh Transformer 场回归、标准 Transformer CFD 对比、适中点数全局交互训练。","典型输入：token 化 (x,fx) / n_heads/n_layers / fun_dim/out_dim / 点数预算。","典型输出：全局注意力预测 / 显存曲线 / checkpoint。","关键配置：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（examples/cfd/EagleMeshTransformer）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：全局自注意力算子基线 / Eagle/Mesh Transformer 场回归 / 标准 Transformer CFD 对比。"].join(`
-`)},scenarios:[{id:"transformer-global-sa",title:"全局自注意力算子基线",summary:"训练标准自注意力作为算子学习基线",fields:[{...e[0],placeholder:"例如：点数适中的网格/点云 CFD 场"},{...e[1],placeholder:"例如：EagleMeshTransformer 配置或 CFD_Benchmark args、n_heads"},e[2],e[3]],capabilitySummary:{current:"全局自注意力算子基线",keyConfig:"n_heads、n_layers、mlp_ratio、dropout、n_hidden",resources:"远程 GPU；严控 NumPoints",deliverables:"checkpoint、相对 L2"},requiredData:["可 token 化数据","点数预算"],expectedOutputs:"基线权重",promptTemplate:["请基于 {{system}} 体系完成 Transformer「全局自注意力算子基线」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transformer（examples/cfd/EagleMeshTransformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 确认点数可承受。","2. 配置注意力超参。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transformer-eagle-mesh-tf",title:"Eagle/Mesh Transformer 场回归",summary:"按 EagleMeshTransformer 示例做网格场回归",fields:[{...e[0],placeholder:"例如：点数适中的网格/点云 CFD 场"},{...e[1],placeholder:"例如：EagleMeshTransformer 配置或 CFD_Benchmark args、n_heads"},e[2],e[3]],capabilitySummary:{current:"Eagle/Mesh Transformer 场回归",keyConfig:"示例配置、数据路径、通道",resources:"GPU",deliverables:"示例跑通日志、预测、权重"},requiredData:["examples/cfd/EagleMeshTransformer 数据"],expectedOutputs:"示例结果",promptTemplate:["请基于 {{system}} 体系完成 Transformer「Eagle/Mesh Transformer 场回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transformer（examples/cfd/EagleMeshTransformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 核对示例目录与依赖。","2. 准备数据。","3. 短跑验证。","4. 正式训练/评估。","5. 记录与 GraphViT 路线差异。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transformer-tf-cfd-bench",title:"标准 Transformer CFD 对比",summary:"在 CFD_Benchmark 协议下与其他注意力/谱模型对比",fields:[{...e[0],placeholder:"例如：点数适中的网格/点云 CFD 场"},{...e[1],placeholder:"例如：EagleMeshTransformer 配置或 CFD_Benchmark args、n_heads"},e[2],e[3]],capabilitySummary:{current:"标准 Transformer CFD 对比",keyConfig:"同协议指标、显存",resources:"同卡对比",deliverables:"对比表"},requiredData:["统一 datapipe"],expectedOutputs:"对比报告",promptTemplate:["请基于 {{system}} 体系完成 Transformer「标准 Transformer CFD 对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transformer（examples/cfd/EagleMeshTransformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 锁定数据。","2. 跑 Transformer。","3. 并列 Factformer/Galerkin/FNO。","4. 总结权衡。","5. 推荐场景。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transformer-mid-token",title:"适中点数全局交互训练",summary:"在安全点数范围内完成全局交互训练并给出扩容上限",fields:[{...e[0],placeholder:"例如：点数适中的网格/点云 CFD 场"},{...e[1],placeholder:"例如：EagleMeshTransformer 配置或 CFD_Benchmark args、n_heads"},e[2],e[3]],capabilitySummary:{current:"适中点数全局交互训练",keyConfig:"NumPoints 上限、batch、梯度检查点可选",resources:"单卡摸上限",deliverables:"可跑点数上限、训练配置"},requiredData:["可降采样数据"],expectedOutputs:"扩容指南",promptTemplate:["请基于 {{system}} 体系完成 Transformer「适中点数全局交互训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transformer（examples/cfd/EagleMeshTransformer）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、space_dim、time_input、unified_pos。","1. 递增点数试跑。","2. 记录 OOM 点。","3. 选定工作点数。","4. 完整训练。","5. 超限建议换窗口/图/谱模型。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},V={id:"transolver",name:"Transolver",category:"onemodel",domain:"fluid",tags:["流体","物理注意力","切片Transformer","几何感知"],blurb:"物理切片注意力 Transformer：把 (x,fx) token 做几何感知全局建模，适非结构场。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/cfd/Transolver-Car-Design",intro:{title:"30 秒看懂模型",subtitle:"非结构或结构 token 均可组织成 (x,fx) 时，用物理状态切片注意力做几何感知场预测",suited:["物理切片注意力场预测","Transolver 汽车外流代理","(x,fx) token 几何感知训练","slice_num 物理切片调参"],notFor:["误用 PyG Data 版 Transolver2D/3D 接口","无法提供 (x,fx)","协议混用"],prepare:["(x,fx[,T])","slice_num","n_heads/n_layers","fun_dim/out_dim/geotype"],obtain:["几何感知预测","切片敏感性","checkpoint"],steps:["确认 Benchmark 接口","设 slice_num","冒烟","训练","汽车外流评估"]},install:{banner:"已为你创建 Transolver 环境配置任务",prompt:["请为 Transolver 创建一个可以直接跑通的环境配置任务。","模型定位：物理切片注意力 Transolver（CFD_Benchmark / examples/cfd/Transolver-Car-Design）。","适合任务：物理切片注意力场预测、Transolver 汽车外流代理、(x,fx) token 几何感知训练、slice_num 物理切片调参。","典型输入：(x,fx[,T]) / slice_num / n_heads/n_layers / fun_dim/out_dim/geotype。","典型输出：几何感知预测 / 切片敏感性 / checkpoint。","关键配置：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、slice_num、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（examples/cfd/Transolver-Car-Design）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：物理切片注意力场预测 / Transolver 汽车外流代理 / (x,fx) token 几何感知训练。"].join(`
-`)},scenarios:[{id:"transolver-physics-slice",title:"物理切片注意力场预测",summary:"用 Transolver_block 物理切片注意力预测 CFD 场",fields:[{...e[0],placeholder:"例如：汽车外流、复杂几何非结构表面/体场"},{...e[1],placeholder:"例如：Transolver-Car-Design 数据、(x,fx)、slice_num、权重"},e[2],e[3]],capabilitySummary:{current:"物理切片注意力场预测",keyConfig:"slice_num、n_heads、n_layers、fun_dim、out_dim",resources:"远程 GPU",deliverables:"checkpoint、相对误差"},requiredData:["(x,fx) 协议数据","完整 args"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 Transolver「物理切片注意力场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transolver（examples/cfd/Transolver-Car-Design）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、slice_num、space_dim、time_input、unified_pos。","1. 确认是 Benchmark (x,fx) 接口而非 PyG 版。","2. 设 slice_num 等。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transolver-car-design",title:"Transolver 汽车外流代理",summary:"按 Transolver-Car-Design 示例做汽车外流代理建模",fields:[{...e[0],placeholder:"例如：汽车外流、复杂几何非结构表面/体场"},{...e[1],placeholder:"例如：Transolver-Car-Design 数据、(x,fx)、slice_num、权重"},e[2],e[3]],capabilitySummary:{current:"Transolver 汽车外流代理",keyConfig:"示例配置、几何点云/网格、通道",resources:"GPU",deliverables:"外流预测、示例指标、权重"},requiredData:["examples/cfd/Transolver-Car-Design 数据"],expectedOutputs:"汽车外流代理结果",promptTemplate:["请基于 {{system}} 体系完成 Transolver「Transolver 汽车外流代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transolver（examples/cfd/Transolver-Car-Design）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、slice_num、space_dim、time_input、unified_pos。","1. 核对示例目录。","2. 准备车身几何与标签。","3. 跑通训练/评估。","4. 检查表面压力/流场误差。","5. 给出设计迭代用法。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transolver-xf-geom",title:"(x,fx) token 几何感知训练",summary:"组织几何坐标与特征 token 做几何感知训练",fields:[{...e[0],placeholder:"例如：汽车外流、复杂几何非结构表面/体场"},{...e[1],placeholder:"例如：Transolver-Car-Design 数据、(x,fx)、slice_num、权重"},e[2],e[3]],capabilitySummary:{current:"(x,fx) token 几何感知训练",keyConfig:"geotype、unified_pos、fun_dim、space_dim",resources:"GPU",deliverables:"训练曲线、几何误差分布"},requiredData:["含坐标的 token 数据"],expectedOutputs:"几何感知模型",promptTemplate:["请基于 {{system}} 体系完成 Transolver「(x,fx) token 几何感知训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transolver（examples/cfd/Transolver-Car-Design）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、slice_num、space_dim、time_input、unified_pos。","1. 检查 x 与 fx 对齐。","2. 训练。","3. 按几何分区误差。","4. 弱几何区加强特征。","5. 固化协议。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"transolver-slice-tune",title:"slice_num 物理切片调参",summary:"扫描物理切片数平衡表达力与算力",fields:[{...e[0],placeholder:"例如：汽车外流、复杂几何非结构表面/体场"},{...e[1],placeholder:"例如：Transolver-Car-Design 数据、(x,fx)、slice_num、权重"},e[2],e[3]],capabilitySummary:{current:"slice_num 物理切片调参",keyConfig:"slice_num 扫描、显存、误差",resources:"短训扫描",deliverables:"slice_num 曲线、推荐值"},requiredData:["验证集"],expectedOutputs:"推荐 slice_num",promptTemplate:["请基于 {{system}} 体系完成 Transolver「slice_num 物理切片调参」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Transolver（examples/cfd/Transolver-Car-Design）执行，关键参数关注：act、dropout、fun_dim、out_dim、geotype、mlp_ratio、n_heads、n_hidden、n_layers、ref、shapelist、slice_num、space_dim、time_input、unified_pos。","1. 设计扫描列表。","2. 短训。","3. 选推荐值。","4. 写入配置。","5. 提醒勿与 PyG 版参数混用。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},J={id:"u_fno",name:"U-FNO",category:"onemodel",domain:"fluid",tags:["流体","U-FNO","多尺度","谱卷积"],blurb:"U-Net 增强的 FNO：谱卷积主干旁加 U 型分支，同时抓全局低频与局部细节。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/U_FNO.py",intro:{title:"30 秒看懂模型",subtitle:"规则网格既要 FNO 全局建模又要加强局部细节时的对比实验主干",suited:["U-Net 增强 FNO 训练","全局谱+局部细节联合预测","U_FNO 多尺度场代理","谱分支与 U 分支消融"],notFor:["无法形成规则网格","非结构仍强依赖 2D Geo 投影","只要纯谱基线"],prepare:["规则网格","modes","fun_dim/out_dim","shapelist"],obtain:["联合预测","分支消融表","checkpoint"],steps:["核对网格","设 modes","双分支冒烟","训练","消融 U 分支"]},install:{banner:"已为你创建 U-FNO 环境配置任务",prompt:["请为 U-FNO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark U-Net enhanced FNO。","适合任务：U-Net 增强 FNO 训练、全局谱+局部细节联合预测、U_FNO 多尺度场代理、谱分支与 U 分支消融。","典型输入：规则网格 / modes / fun_dim/out_dim / shapelist。","典型输出：联合预测 / 分支消融表 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/U_FNO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：U-Net 增强 FNO 训练 / 全局谱+局部细节联合预测 / U_FNO 多尺度场代理。"].join(`
-`)},scenarios:[{id:"u_fno-ufno-train",title:"U-Net 增强 FNO 训练",summary:"训练谱主干+U 型分支的 U_FNO",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"U-Net 增强 FNO 训练",keyConfig:"modes、n_hidden、fun_dim、out_dim、shapelist",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["规则网格数据"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 U-FNO「U-Net 增强 FNO 训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-FNO（src/onescience/models/cfd_benchmark/U_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、time_input、unified_pos。","1. 验证 shapelist。","2. 配置 modes。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_fno-spec-local",title:"全局谱+局部细节联合预测",summary:"评估全局低频与局部细节的联合预测效果",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"全局谱+局部细节联合预测",keyConfig:"反归一化、局部误差切片、modes",resources:"单卡评估",deliverables:"全局/局部误差分解"},requiredData:["权重","测试集"],expectedOutputs:"误差分解报告",promptTemplate:["请基于 {{system}} 体系完成 U-FNO「全局谱+局部细节联合预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-FNO（src/onescience/models/cfd_benchmark/U_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、time_input、unified_pos。","1. 推理。","2. 分解低频/高频误差。","3. 对照纯 FNO。","4. 结论 U 分支收益。","5. 建议是否保留。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_fno-ufno-surr",title:"U_FNO 多尺度场代理",summary:"部署为多尺度场代理做推理服务准备",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"U_FNO 多尺度场代理",keyConfig:"checkpoint、输入协议、输出通道",resources:"推理 GPU",deliverables:"代理场、延迟、接口说明"},requiredData:["权重","样例输入"],expectedOutputs:"代理包",promptTemplate:["请基于 {{system}} 体系完成 U-FNO「U_FNO 多尺度场代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-FNO（src/onescience/models/cfd_benchmark/U_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、time_input、unified_pos。","1. 加载。","2. 批量推理。","3. 测延迟。","4. 固化预处理。","5. 给出调用示例。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_fno-branch-ablate",title:"谱分支与 U 分支消融",summary:"消融 U 分支或谱分支量化各自贡献",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"谱分支与 U 分支消融",keyConfig:"分支开关、同种子、指标",resources:"短训/评估",deliverables:"消融表"},requiredData:["可消融实现或对照 FNO"],expectedOutputs:"消融结论",promptTemplate:["请基于 {{system}} 体系完成 U-FNO「谱分支与 U 分支消融」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-FNO（src/onescience/models/cfd_benchmark/U_FNO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、time_input、unified_pos。","1. 跑满配 U_FNO。","2. 对照 FNO。","3. 比较局部指标。","4. 写消融结论。","5. 指导最终结构。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},X={id:"u_net",name:"U-Net",category:"onemodel",domain:"fluid",tags:["流体","U-Net","多尺度","场代理"],blurb:"CFD_Benchmark U-shape 场代理：统一 (x,fx) 接口进多尺度编码解码，非图像 UNet。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/U_Net.py",intro:{title:"30 秒看懂模型",subtitle:"可稳定 reshape 为规则网格时，用 U-shape 主干做场代理并与 FNO 族同协议对比",suited:["U-shape 网格场代理训练","Benchmark U-Net 多尺度回归","规则网格编码解码预测","与 FNO/U_FNO 同协议对比"],notFor:["直接喂 BCHW 图像 batch","硬 reshape 非结构点云","与 DeepCFD 图像接口混淆"],prepare:["可 reshape 网格","fun_dim/out_dim/task","modes/n_hidden","shapelist"],obtain:["U-shape 预测","同协议对比表","checkpoint"],steps:["确认非图像接口","设 task/shapelist","冒烟","训练","与 FNO 对比"]},install:{banner:"已为你创建 U-Net 环境配置任务",prompt:["请为 U-Net 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark U-shape field surrogate（非 DeepCFD 图像 UNet）。","适合任务：U-shape 网格场代理训练、Benchmark U-Net 多尺度回归、规则网格编码解码预测、与 FNO/U_FNO 同协议对比。","典型输入：可 reshape 网格 / fun_dim/out_dim/task / modes/n_hidden / shapelist。","典型输出：U-shape 预测 / 同协议对比表 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/U_Net.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：U-shape 网格场代理训练 / Benchmark U-Net 多尺度回归 / 规则网格编码解码预测。"].join(`
-`)},scenarios:[{id:"u_net-ushape-surr",title:"U-shape 网格场代理训练",summary:"按 Benchmark (x,fx) 接口训练 U-shape 场代理",fields:[{...e[0],placeholder:"例如：可 reshape 的规则网格 CFD 场"},{...e[1],placeholder:"例如：CFD_Benchmark (x,fx) 配置、task、shapelist、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"U-shape 网格场代理训练",keyConfig:"task、fun_dim、out_dim、n_hidden、shapelist",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["可 reshape 数据","非 BCHW 图像接口"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 U-Net「U-shape 网格场代理训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-Net（src/onescience/models/cfd_benchmark/U_Net.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 拒绝误用图像 UNet 接口。","2. 配置 task 与通道。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_net-bench-unet",title:"Benchmark U-Net 多尺度回归",summary:"多尺度编码解码回归目标场并检查跳跃连接收益",fields:[{...e[0],placeholder:"例如：可 reshape 的规则网格 CFD 场"},{...e[1],placeholder:"例如：CFD_Benchmark (x,fx) 配置、task、shapelist、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"Benchmark U-Net 多尺度回归",keyConfig:"深度、n_hidden、modes（若用）、归一化",resources:"GPU",deliverables:"训练曲线、多尺度误差"},requiredData:["标签场"],expectedOutputs:"回归结果",promptTemplate:["请基于 {{system}} 体系完成 U-Net「Benchmark U-Net 多尺度回归」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-Net（src/onescience/models/cfd_benchmark/U_Net.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 训练。","2. 分层可视化。","3. 评估。","4. 细节不足则加宽。","5. 记录配置。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_net-codec-pred",title:"规则网格编码解码预测",summary:"对规则网格新样本做编码解码预测",fields:[{...e[0],placeholder:"例如：可 reshape 的规则网格 CFD 场"},{...e[1],placeholder:"例如：CFD_Benchmark (x,fx) 配置、task、shapelist、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"规则网格编码解码预测",keyConfig:"checkpoint、shapelist、反归一化",resources:"单卡",deliverables:"预测场、误差图"},requiredData:["权重","测试网格"],expectedOutputs:"预测",promptTemplate:["请基于 {{system}} 体系完成 U-Net「规则网格编码解码预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-Net（src/onescience/models/cfd_benchmark/U_Net.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 核对 shapelist。","2. 推理。","3. 评估。","4. 失败查 reshape。","5. 输出可视化。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_net-vs-fno-family",title:"与 FNO/U_FNO 同协议对比",summary:"同一 datapipe 下对比 U-Net 与谱算子族",fields:[{...e[0],placeholder:"例如：可 reshape 的规则网格 CFD 场"},{...e[1],placeholder:"例如：CFD_Benchmark (x,fx) 配置、task、shapelist、fun_dim/out_dim"},e[2],e[3]],capabilitySummary:{current:"与 FNO/U_FNO 同协议对比",keyConfig:"同划分同指标",resources:"同协议对比",deliverables:"对比表、场景推荐"},requiredData:["三模型配置"],expectedOutputs:"对比报告",promptTemplate:["请基于 {{system}} 体系完成 U-Net「与 FNO/U_FNO 同协议对比」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-Net（src/onescience/models/cfd_benchmark/U_Net.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 锁定协议。","2. 评估三者。","3. 比较全局/局部误差。","4. 给出推荐。","5. 避免接口混用说明。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},Q={id:"u_no",name:"U-NO",category:"onemodel",domain:"fluid",tags:["流体","U型算子","FNO","多尺度"],blurb:"U 型神经算子：下/上采样路径注入谱算子，比 FNO 更强的多尺度算子学习。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/src/onescience/models/cfd_benchmark/U_NO.py",intro:{title:"30 秒看懂模型",subtitle:"需要比 FNO 更强多尺度表达又保持算子学习属性的规则网格 CFD 任务",suited:["U 型神经算子训练","多尺度谱路径场预测","U_NO 下采样稳定性校验","强于 FNO 的多尺度代理"],notFor:["结构尺寸过小多次下采样不稳","无正确 shapelist","只要最轻 FNO"],prepare:["足够大的规则网格","modes、task","fun_dim/out_dim","shapelist"],obtain:["多尺度算子预测","下采样诊断","checkpoint"],steps:["检查尺寸可下采样","设 modes/task","冒烟","训练","对比 FNO"]},install:{banner:"已为你创建 U-NO 环境配置任务",prompt:["请为 U-NO 创建一个可以直接跑通的环境配置任务。","模型定位：CFD_Benchmark U-shaped Neural Operator。","适合任务：U 型神经算子训练、多尺度谱路径场预测、U_NO 下采样稳定性校验、强于 FNO 的多尺度代理。","典型输入：足够大的规则网格 / modes、task / fun_dim/out_dim / shapelist。","典型输出：多尺度算子预测 / 下采样诊断 / checkpoint。","关键配置：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience 入口（src/onescience/models/cfd_benchmark/U_NO.py）与数据/权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令（含假输入 forward / shape 检查），并说明下一步可以进入：U 型神经算子训练 / 多尺度谱路径场预测 / U_NO 下采样稳定性校验。"].join(`
-`)},scenarios:[{id:"u_no-uno-train",title:"U 型神经算子训练",summary:"训练带谱算子的 U 型多尺度神经算子",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"U 型神经算子训练",keyConfig:"modes、n_hidden、task、fun_dim、out_dim、shapelist",resources:"远程 GPU",deliverables:"checkpoint、相对 L2"},requiredData:["足够分辨率网格"],expectedOutputs:"权重",promptTemplate:["请基于 {{system}} 体系完成 U-NO「U 型神经算子训练」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-NO（src/onescience/models/cfd_benchmark/U_NO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 确认尺寸支撑下采样。","2. 配置 modes/task。","3. 冒烟。","4. 训练。","5. 存盘。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_no-ms-spec-path",title:"多尺度谱路径场预测",summary:"在多尺度路径上注入谱卷积做场预测",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"多尺度谱路径场预测",keyConfig:"各尺度 modes、跳跃、反归一化",resources:"GPU",deliverables:"多尺度预测、尺度误差"},requiredData:["权重或训练配置"],expectedOutputs:"预测与尺度分析",promptTemplate:["请基于 {{system}} 体系完成 U-NO「多尺度谱路径场预测」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-NO（src/onescience/models/cfd_benchmark/U_NO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 推理。","2. 按尺度分析误差。","3. 对照 FNO。","4. 调整 modes。","5. 固化。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_no-down-stable",title:"U_NO 下采样稳定性校验",summary:"校验网格尺寸与下采样次数匹配，避免过小特征图",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"U_NO 下采样稳定性校验",keyConfig:"shapelist、下采样级数、最小边长",resources:"本地诊断+短冒烟",deliverables:"稳定性检查表、安全配置"},requiredData:["目标分辨率"],expectedOutputs:"可训配置",promptTemplate:["请基于 {{system}} 体系完成 U-NO「U_NO 下采样稳定性校验」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-NO（src/onescience/models/cfd_benchmark/U_NO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 计算每级尺寸。","2. 过小则减层或升分辨率。","3. 冒烟。","4. 再训练。","5. 写入约束。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"u_no-beyond-fno",title:"强于 FNO 的多尺度代理",summary:"证明在多尺度场景相对 FNO 的增益并形成代理",fields:[{...e[0],placeholder:"例如：Darcy 流、圆柱绕流、管道稳态场"},{...e[1],placeholder:"例如：CFD_Benchmark datapipe、fun_dim/out_dim、geotype、shapelist、modes"},e[2],e[3]],capabilitySummary:{current:"强于 FNO 的多尺度代理",keyConfig:"同数据对比、局部指标、算力",resources:"同卡对比",deliverables:"增益报告、代理部署说明"},requiredData:["FNO 与 U_NO 结果"],expectedOutputs:"选型与部署建议",promptTemplate:["请基于 {{system}} 体系完成 U-NO「强于 FNO 的多尺度代理」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience U-NO（src/onescience/models/cfd_benchmark/U_NO.py）执行，关键参数关注：act、fun_dim、out_dim、geotype、modes、n_hidden、ref、shapelist、space_dim、task、time_input、unified_pos。","1. 对齐评估。","2. 量化增益。","3. 算力权衡。","4. 决定上线 U_NO 或回退 FNO。","5. 输出操作说明。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},K=[v,N,F,T,A,x,G,U,B,E,w,j,L,R,q,I,H,z,W,V,J,X,Q],Y={id:"matris",name:"MatRIS",category:"onemodel",domain:"materials",tags:["材料化学","图神经势","能量力应力","磁矩"],blurb:"面向晶体与分子的图神经势函数，预测能量、力、应力与磁矩，可接入 ASE 弛豫与 MD。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/matris",intro:{title:"30 秒看懂模型",subtitle:"在晶体或分子结构上预测能量、力、应力和磁矩，并可经 MatRISCalculator / StructOptimizer 做弛豫与 MD",suited:["E/F/S 预测","磁矩读出","结构弛豫","ASE 分子动力学"],notFor:["只做能带/DOS","元素超出覆盖","无可用预训练权重"],prepare:["CIF/POSCAR/Structure","task=e/ef/efs/efsm","matris_10m_oam 或 mp"],obtain:["e/f/s/m 字典","弛豫结构与轨迹","ASE 能量/力/应力"],steps:["选择 task 与权重","读取结构并构图","小结构 smoke test","正式推理或弛豫","接入 ASE calculator/MD"]},install:{banner:"已为你创建 MatRIS 环境配置任务",prompt:["请为 MatRIS 创建一个可以直接跑通的环境配置任务。","模型定位：Material Representation Learning with Interatomic Structure，图神经材料势函数（atom graph + line graph）。","适合任务：E/F/S/M 预测、磁矩评估、StructOptimizer 结构弛豫、MatRISCalculator + ASE MD。","典型输入：CIF/POSCAR/pymatgen Structure/ASE Atoms、task（e/ef/efs/efsm）、matris_10m_oam 或 matris_10m_mp、ONESCIENCE_MODELS_DIR。","典型输出：forward 字典 e/f/s/m、弛豫轨迹、ASE get_potential_energy/forces/stress/magnetic_moments。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience MatRIS 示例（examples/matchem/matris，含 cif_file/demo.cif、test_relaxation.py）与权重路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式（优先 ONESCIENCE_MODELS_DIR/matris/）。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：预测能量力应力、评估原子磁矩、ASE 结构弛豫、ASE 短时程 MD。"].join(`
-`)},scenarios:[{id:"matris-predict-efs",title:"预测能量力应力",summary:"对晶体/分子结构做 MatRIS forward，输出能量、力和可选应力",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：Si 立方胞、Cu 体相、合金超胞"},{id:"inputData",label:"结构与模型",placeholder:"例如：cif_file/demo.cif 或 POSCAR + matris_10m_oam",multiline:!0},{id:"goal",label:"预测目标",placeholder:"例如：task=efs 输出每原子能、力与应力，核对 is_intensive",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 单结构 CPU 冒烟"}],capabilitySummary:{current:"MatRIS 能量/力/应力推理",keyConfig:"task=e/ef/efs、pairwise_cutoff=6、three_body_cutoff=4、is_conservation、is_intensive、matris_10m_oam/mp",resources:"小结构可 CPU；预训练 efsm/efs 建议 GPU；三体 line graph 显存随 cutoff 增长",deliverables:"e/f/s 字段、shape/单位检查报告、是否可进入弛豫的结论"},requiredData:["CIF/POSCAR 或 pymatgen.Structure / ase.Atoms","task 字符串（e / ef / efs）与是否守恒力 is_conservation","权重：matris_10m_oam、matris_10m_mp 或本地 MatRIS_10M_*.pth.tar"],expectedOutputs:"预测字典 e/f/s、atoms_per_graph、单位与 is_intensive 核对说明",promptTemplate:["请基于 {{system}} 体系完成 MatRIS「预测能量力应力」任务。","结构与模型：{{inputData}}","预测目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MatRIS 流程执行：","1. 读取结构为 pymatgen.Structure 或 ase.Atoms，确认元素、PBC、cell 与原子间距合理。","2. 用 MatRIS.load(model_name='matris_10m_oam'|'matris_10m_mp') 或本地 checkpoint；离线优先 ONESCIENCE_MODELS_DIR/matris/。","3. 经 model.graph_converter(structure) 得到 RadiusGraph，调用 model([graph], task='ef'|'efs', is_training=False)。","4. 核对输出 e（is_intensive 时为每原子能）、f 形状 (N,3)、s 为 (3,3) 且注意内部 GPa 与 ASE eV/ų 换算。","5. 输出字段检查报告，并说明下一步可做磁矩（efsm）、StructOptimizer 弛豫或 MatRISCalculator MD。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"matris-magmom",title:"评估原子磁矩",summary:"用 task 含 m 的 MagmomHead 预测逐原子磁矩并做异常检查",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：磁性合金、过渡金属氧化物"},{id:"inputData",label:"结构与模型",placeholder:"例如：demo.cif + matris_10m_oam，task=efsm 或 em",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：输出逐原子磁矩统计，定位异常原子",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU"}],capabilitySummary:{current:"MatRIS 磁矩预测（MagmomHead）",keyConfig:"task=em/efsm、MagmomHead、元素覆盖（max 94）、预训练 matris_10m_oam/mp",resources:"efsm 建议 GPU；大三体图注意显存",deliverables:"逐原子磁矩 m、统计量与异常原子列表"},requiredData:["含磁性相关原子的结构文件","task 含 m（em 或 efsm）","matris_10m_oam / matris_10m_mp 或兼容本地权重"],expectedOutputs:"m 列表、均值/极值统计、异常原子索引与可用性说明",promptTemplate:["请基于 {{system}} 体系完成 MatRIS「评估原子磁矩」任务。","结构与模型：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MatRIS 流程执行：","1. 确认体系需要磁矩读出，设置 task='em' 或 'efsm'（MagmomHead 仅在 task 含 m 时计算）。","2. 加载预训练权重并构图；可用 MatRISCalculator(..., task='efsm') 经 ASE get_magnetic_moments() 校验。","3. 核对 m 形状为每结构 (N,)，与原子顺序一致；说明该输出不等于自洽电子结构求解。","4. 统计磁矩分布，标记异常近邻或空邻居导致的可疑原子。","5. 给出能否用于磁性相关筛选/弛豫的结论，并提示下一步可接 StructOptimizer(task='efsm')。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"matris-relax",title:"ASE 结构弛豫",summary:"用 StructOptimizer / MatRISCalculator 做原子与可选晶胞弛豫",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：无机晶体、合金超胞"},{id:"inputData",label:"初始结构与模型",placeholder:"例如：examples/matchem/matris/cif_file/demo.cif + matris_10m_oam",multiline:!0},{id:"goal",label:"弛豫目标",placeholder:"例如：FIRE + FrechetCellFilter，fmax=0.05，steps≤500",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU；先 10 step 冒烟"}],capabilitySummary:{current:"MatRIS StructOptimizer 结构弛豫",keyConfig:"task=ef/efs/efsm、optimizer=FIRE/BFGS、fmax、steps、relax_cell、ase_filter=FrechetCellFilter",resources:"GPU；先短步数 smoke test 再正式收敛",deliverables:"final_structure、trajectory（能量/力/应力/磁矩）、收敛报告"},requiredData:["初始 CIF/POSCAR（参考 cif_file/demo.cif）","模型名 matris_10m_oam 或 mp，以及 task（优化通常 ef/efs/efsm）","收敛判据 fmax、最大步数、是否 relax_cell"],expectedOutputs:"优化结构、轨迹能量/最大力、收敛与否说明",promptTemplate:["请基于 {{system}} 体系完成 MatRIS「ASE 结构弛豫」任务。","初始结构与模型：{{inputData}}","弛豫目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MatRIS 示例（test_relaxation.py）执行：","1. 用 StructOptimizer(model=..., task='ef'|'efs'|'efsm', optimizer='FIRE', device=...) 封装。","2. 从 CIF/POSCAR 读入 Structure/Atoms；先做约 10 step 短弛豫确认能量下降与最大力合理。","3. 正式运行 relax(atoms=..., steps=..., fmax=..., relax_cell=..., ase_filter='FrechetCellFilter')。","4. 记录 trajectory 末帧能量、力、应力（及磁矩），检查异常近邻与元素覆盖。","5. 输出最终结构路径与是否可进入 ASE MD 的结论。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"matris-ase-md",title:"ASE 短时程 MD",summary:"把 MatRISCalculator 接到 ASE，做短时程 NVE/NVT 稳定性检查",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：Cu 体相、分子材料"},{id:"inputData",label:"模型与初始结构",placeholder:"例如：matris_10m_oam + 弛豫后的结构 / POSCAR",multiline:!0},{id:"goal",label:"MD 目标",placeholder:"例如：300K NVT 短时程，检查能量漂移与异常力",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU"}],capabilitySummary:{current:"MatRISCalculator + ASE 分子动力学",keyConfig:"MatRISCalculator(task='efs'|'efsm')、温度、步长、系综、短时程 smoke test",resources:"GPU；先短程 NVE/NVT 再拉长",deliverables:"MD 轨迹、能量/温度曲线、稳定性结论与脚本"},requiredData:["可用 MatRIS 权重与 MatRISCalculator 配置","已预优化或合理的初始结构","温度/步长/系综与输出间隔"],expectedOutputs:"轨迹文件、能量漂移检查、是否可拉长 MD 的建议",promptTemplate:["请基于 {{system}} 体系完成 MatRIS「ASE 短时程 MD」任务。","模型与初始结构：{{inputData}}","MD 目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MatRIS / ASE 流程执行：","1. 绑定 atoms.calc = MatRISCalculator(model=..., task='efs'|'efsm', device=...)，先单点核对能量与力 shape。","2. 确认非周期结构 cell 已被 calculator 合理扩展；周期结构核对 PBC。","3. 先做短时程 NVE/NVT smoke test，检查能量漂移、异常力与近邻爆炸。","4. 与同一构型上的 forward/Calculator 结果做一致性抽查（总能 vs is_intensive 换算）。","5. 给出正式拉长 MD 的参数建议与后处理入口。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},Z={id:"uma_escn_md",name:"UMA-eSCN-MD",category:"onemodel",domain:"materials",tags:["材料化学","通用势","等变backbone","EFS"],blurb:"UMA eSCNMD 等变主干，支持 charge/spin/dataset 条件，用于 E/F/S 推理、微调与弛豫。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/uma",intro:{title:"30 秒看懂模型",subtitle:"把原子结构编码为 SO(3) 等变节点表示，配合 UMA head 做能量/力/应力推理、微调与弛豫/MD",suited:["带条件 EFS 推理","backbone 微调","误差评估","弛豫与 MD"],notFor:["无原子构型","缺 dataset/charge/spin 契约","应优先上 MoE 的强多域冲突"],prepare:["ASE-LMDB/AtomicData","uma-s-1p1_converted.pt","Jd.pt 与 elem_refs"],obtain:["energy/forces/stress","微调 checkpoint","弛豫/MD 轨迹"],steps:["核对字段与 PBC","装配 escnmd_backbone+head","小 batch smoke","微调或评估","inference 弛豫/MD"]},install:{banner:"已为你创建 UMA-eSCN-MD 环境配置任务",prompt:["请为 UMA-eSCN-MD 创建一个可以直接跑通的环境配置任务。","模型定位：UMA HydraModel 的 eSCNMDBackbone（注册名 escnmd_backbone），等变通用材料势主干。","适合任务：带 dataset/charge/spin 的 E/F/S 推理、Hydra 微调、测试集评估、FAIRChemCalculator 弛豫/MD。","典型输入：pos/atomic_numbers/cell/pbc/charge/spin/dataset、ASE-LMDB、uma-s-1p1_converted.pt、Jd.pt、elem_refs/normalizer_rmsd。","典型输出：MLP_Energy_Head / MLP_EFS_Head 预测、微调 checkpoint、energy/force MAE、弛豫结构与 MD 轨迹。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience UMA demo（examples/matchem/uma/demo）与 inference/、weight/Jd.pt、checkpoint 路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式（转换后 checkpoint，勿混用未转换 uma-s-1p1.pt）。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：带条件的 EFS 推理、微调 eSCNMD backbone、评估能量力应力误差、结构弛豫与分子动力学。"].join(`
-`)},scenarios:[{id:"uma_escn_md-efs-infer",title:"带条件的 EFS 推理",summary:"按 dataset/charge/spin 与 PBC 配置，用 escnmd_backbone + head 做 E/F/S 推理",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：OC20 吸附、OMAT 晶体、OMOL 分子"},{id:"inputData",label:"结构与条件字段",placeholder:"例如：ASE Atoms/AtomicData + dataset=omat + charge/spin + uma-s-1p1_converted.pt",multiline:!0},{id:"goal",label:"推理目标",placeholder:"例如：输出 energy/forces，可选 stress（regress_stress）",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU；单 batch 冒烟"}],capabilitySummary:{current:"UMA eSCNMD 条件化 E/F/S 推理",keyConfig:"otf_graph=true、cutoff=5.0、max_neighbors=300、dataset_list、always_use_pbc、regress_stress、jd_path、MLP_EFS_Head",resources:"GPU；显存随边数与 max_neighbors 增长",deliverables:"energy/forces/stress、字段检查与 PBC/构图日志"},requiredData:["pos、atomic_numbers、batch、natoms、cell、pbc、charge、spin、dataset","非空 dataset_list，且输入 dataset 被覆盖（如 omat/oc20/omol）","转换后 checkpoint 与 Jd.pt（ONESCIENCE_MODELS_DIR/UMA/ 或 examples/matchem/uma/weight/）"],expectedOutputs:"E/F/S 预测、shape 检查、异常无边样本列表",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MD「带条件的 EFS 推理」任务。","结构与条件字段：{{inputData}}","推理目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience UMA 流程执行：","1. 确认 AtomicData 字段齐全；分子设 always_use_pbc=false 并给足够大真空盒；晶体/吸附允许 PBC；同一 batch 不混合全 true/全 false。","2. 装配 HydraModel：backbone.model=escnmd_backbone，otf_graph=true，cutoff=5.0，max_neighbors=300，显式 jd_path。","3. 按目标选择 head：energy-only 用 MLP_Energy_Head；EF 用 MLP_EFS_Head 且 regress_stress=false；EFS 开 regress_stress=true。","4. 小 batch forward smoke test：检查构图有边、输出 dict 与 head 约定一致。","5. 正式推理后报告结果路径，并说明可进入评估、微调或 inference/relax_*.py、run_molecular_md.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_md-finetune",title:"微调 eSCNMD backbone",summary:"在目标 ASE-LMDB 上用 Hydra demo 微调 escnmd_backbone",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：OC20 S2EF、OMAT 无机材料"},{id:"inputData",label:"数据与 checkpoint",placeholder:"例如：data/oc20_finetune train/val ASE-LMDB + uma-s-1p1_converted.pt",multiline:!0},{id:"goal",label:"微调目标",placeholder:"例如：regression-tasks=ef，先 1 epoch 冒烟再正式训练",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程多卡 GPU / demo YAML + dry-run"}],capabilitySummary:{current:"Hydra 微调 UMA eSCNMD backbone",keyConfig:"checkpoint_location、elem_refs、normalizer_rmsd、lr=0.0004、weight_decay=0.001、batch_size、evaluate_every_n_steps",resources:"建议远程 GPU/DCU；先 run.sh --dry-run 与单卡冒烟",deliverables:"微调 checkpoint、训练曲线、energy/force MAE"},requiredData:["ASE-LMDB train/val（可用 scripts/create_uma_finetune_dataset.py 从 extxyz 转换）","与数据同源的 elem_refs、normalizer_rmsd、dataset_name、heads、tasks_list","uma-s-1p1_converted.pt 与 Jd.pt"],expectedOutputs:"checkpoint、lcurve/日志、验证 MAE 与下一步评估入口",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MD「微调 eSCNMD backbone」任务。","数据与 checkpoint：{{inputData}}","微调目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience examples/matchem/uma 流程执行：","1. 用 scripts/create_uma_finetune_dataset.py 准备 ASE-LMDB，并生成/写入 elem_refs 与 normalizer_rmsd（可用 update_demo_config.py）。","2. 配置 Hydra：backbone.model=escnmd_backbone，checkpoint_location 指向转换后权重，jd_path 显式设置。","3. 对齐 head/task：EF 勿误开 regress_stress；EFS 同步 stress 标签与 loss。","4. 先 bash demo/run.sh --config demo/configs/oc20_ef_*.yaml --dry-run，再短 epoch/max_steps 冒烟确认 loss 与 checkpoint。","5. 正式训练后输出验证误差与 checkpoint 路径，说明可进入误差评估或弛豫/MD。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_md-evaluate",title:"评估能量力应力误差",summary:"在独立测试集上计算 energy/force/stress MAE 并判断能否进 MD",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：OC20 val_id、自建测试集"},{id:"inputData",label:"模型与测试数据",placeholder:"例如：finetune checkpoint + val.aselmdb，字段与训练一致",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：报告 E/F/S MAE，定位异常高能/无力近邻样本",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU"}],capabilitySummary:{current:"评估 UMA eSCNMD 势精度",keyConfig:"energy/force/stress MAE、dataset_name 同源、elem_refs/normalizer、PBC 分流",resources:"GPU；大批量注意边数与 batch",deliverables:"误差表、异常样本列表、能否弛豫/MD 的结论"},requiredData:["待评估 checkpoint 与训练时一致的 head/task 配置","独立测试 ASE-LMDB/AtomicData（标签单位与训练一致）","elem_refs、normalizer_rmsd、dataset 映射"],expectedOutputs:"E/F/S 误差表、异常结构列表、部署建议",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MD「评估能量力应力误差」任务。","模型与测试数据：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience UMA 评估约定执行：","1. 确认测试集字段、单位、dataset、elem_refs、normalizer_rmsd 与训练同源，避免统计量错配。","2. 在独立测试集上计算 energy/force（及 stress）MAE/RMSE，必要时按 dataset 或体系分组。","3. 定位最大误差帧、无边图、PBC 断言失败或 charge/spin 缺失样本。","4. 给出明确结论：能否进入 FAIRChemCalculator 弛豫/MD，或需先修复数据/继续微调。","5. 若发现配置不一致，先暂停并给出最小修复方案。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_md-relax-md",title:"结构弛豫与分子动力学",summary:"用 FAIRChemCalculator 做晶体弛豫或分子短 MD",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：Fe 体相弛豫、H2O 分子 MD"},{id:"inputData",label:"模型与初始结构",placeholder:"例如：checkpoint/uma-s-1p1.pt + POSCAR/分子；task_name=omat 或 omol",multiline:!0},{id:"goal",label:"模拟目标",placeholder:"例如：LBFGS+FrechetCellFilter fmax=0.05，或 400K Langevin 短 MD",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU"}],capabilitySummary:{current:"UMA ASE 弛豫 / 分子 MD",keyConfig:"load_predict_unit、FAIRChemCalculator(task_name)、LBFGS/FrechetCellFilter、Langevin 温度与步长",resources:"GPU；先短弛豫/短轨迹再拉长",deliverables:"优化 CIF/traj、MD 轨迹与能量稳定性报告"},requiredData:["可用 UMA checkpoint 与 Jd 环境（ONESCIENCE_UMA_JD_PATH）","初始结构；task_name 与域匹配（omat/omol/oc20 等）","弛豫判据或 MD 温度/步长/摩擦"],expectedOutputs:"Fe_optimized.cif/traj 或 my_md.traj、能量/力检查、拉长计划",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MD「结构弛豫与分子动力学」任务。","模型与初始结构：{{inputData}}","模拟目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience examples/matchem/uma/inference 执行：","1. load_predict_unit(ckpt) + FAIRChemCalculator(predictor, task_name=...)；晶体参考 relax_inorganic_crystal.py，分子参考 run_molecular_md.py。","2. 弛豫：FrechetCellFilter + LBFGS，先短步数确认 fmax 下降；MD：MaxwellBoltzmann + Langevin，先短时程检查能量漂移。","3. 核对 task_name 与体系域一致，分子注意真空盒与非周期设定。","4. 保存结构/轨迹（cif、traj、forces），记录最终能量与最大力。","5. 给出是否可拉长采样的结论与后处理入口。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},$={id:"uma_escn_moe",name:"UMA-eSCN-MoE",category:"onemodel",domain:"materials",tags:["材料化学","MoE","多域势","专家路由"],blurb:"UMA MoE 多域材料势主干，按 dataset/电荷/自旋路由专家，适合多域微调与目标域部署。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/uma",intro:{title:"30 秒看懂模型",subtitle:"在 eSCNMD 等变表征上加入 MOLE 专家路由，按 charge/spin/dataset（可选组成）混合专家，服务多域 E/F/S",suited:["MoE 推理","多域联合微调","专家路由评估","合并部署目标域"],notFor:["单域已够用","缺路由字段","用 MoE 掩盖标签/PBC 错误"],prepare:["多域 ASE-LMDB","escnmd_moe_backbone 配置","num_experts 与 Jd.pt"],obtain:["E/F/S 与专家系数","分域误差报告","merge 后普通 backbone"],steps:["先跑普通 UMA baseline","启用 MoE 与路由","监控专家利用率","分域评估","merge_MOLE_model 部署"]},install:{banner:"已为你创建 UMA-eSCN-MoE 环境配置任务",prompt:["请为 UMA-eSCN-MoE 创建一个可以直接跑通的环境配置任务。","模型定位：eSCNMDMoeBackbone（注册名 escnmd_moe_backbone），在 UMA eSCNMD 上叠加 MOLE/MoE 专家路由。","适合任务：多域 MoE 推理、多域联合微调、专家路由评估、merge_MOLE_model 固定上下文部署。","典型输入：多 dataset 的 AtomicData、num_experts/moe_type/layers_moe、uma-s-1p1_converted.pt、Jd.pt、elem_refs/normalizer_rmsd。","典型输出：E/F/S、expert_mixing_coefficients、分域 MAE、合并后的 eSCNMDBackbone。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience UMA 路径（examples/matchem/uma/demo）与 MoE Hydra override、权重与 Jd.pt 是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式；确认普通 escnmd_backbone baseline 已可跑通再上 MoE。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：MoE 势函数推理、多域联合微调、专家路由评估、合并部署目标域。"].join(`
-`)},scenarios:[{id:"uma_escn_moe-infer",title:"MoE 势函数推理",summary:"加载 escnmd_moe_backbone，按体系条件输出 E/F/S 与专家混合系数",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：跨 OMAT/OMOL 的异构结构 batch"},{id:"inputData",label:"结构与 MoE 配置",placeholder:"例如：AtomicData + dataset_list + num_experts=8 + moe_type=so2 + converted checkpoint",multiline:!0},{id:"goal",label:"推理目标",placeholder:"例如：输出 E/F 并记录 expert_mixing_coefficients",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU；小 batch 冒烟"}],capabilitySummary:{current:"UMA MoE backbone 推理",keyConfig:"escnmd_moe_backbone、num_experts=8、moe_type=so2、moe_dropout、use_composition_embedding、cutoff=5.0、max_neighbors=300",resources:"GPU；显存高于普通 eSCNMD，优先小 batch",deliverables:"E/F/S、专家系数、构图与路由检查日志"},requiredData:["与 uma_escn_md 相同的 AtomicData 字段，且 dataset/charge/spin 稳定","覆盖全部推理域的 dataset_list","MoE checkpoint 或可加载的 converted UMA 权重 + Jd.pt"],expectedOutputs:"预测结果、expert_mixing_coefficients、无边/路由异常样本",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MoE「MoE 势函数推理」任务。","结构与 MoE 配置：{{inputData}}","推理目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience UMA MoE 流程执行：","1. 确认普通 escnmd_backbone 对同数据可 forward；再切 backbone.model=escnmd_moe_backbone。","2. 设置 num_experts（常用 4/8）、moe_type=so2、moe_layer_type=pytorch、显式 jd_path；按需 use_composition_embedding。","3. 小 batch smoke：检查 expert_mixing_coefficients、mole_sizes、head 输出与 loss 可算。","4. 正式推理并保存分体系专家系数，排查 dataset 未覆盖或 PBC 混合问题。","5. 说明下一步可做多域微调、路由评估或 merge_MOLE_model。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_moe-multidomain",title:"多域联合微调",summary:"在多 dataset 上联合微调 MoE backbone，缓解域间误差冲突",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：OC20+OMAT 或多任务材料域"},{id:"inputData",label:"多域数据与权重",placeholder:"例如：各域 ASE-LMDB + uma-s-1p1_converted.pt + 分域 elem_refs",multiline:!0},{id:"goal",label:"微调目标",placeholder:"例如：num_experts=8、部分 layers_moe，提升弱域 MAE",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程多卡 GPU"}],capabilitySummary:{current:"多域 MoE 联合微调",keyConfig:"num_experts、layers_moe、moe_type=so2、lr=0.0004、DatasetSpecificMoEWrapper（可选）、elem_refs 同源",resources:"多卡 GPU；先 baseline 再加专家；控制专家数防过拟合",deliverables:"MoE checkpoint、分域验证曲线、专家利用率日志"},requiredData:["多域 train/val ASE-LMDB 与完整 dataset_list","各域或统一约定的 elem_refs、normalizer_rmsd、heads/tasks_list","转换后 UMA checkpoint 与 Jd.pt"],expectedOutputs:"多域 checkpoint、分 dataset MAE、训练期 MOLE stats",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MoE「多域联合微调」任务。","多域数据与权重：{{inputData}}","微调目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience UMA MoE 规划执行：","1. 先建立普通 uma_escn_md baseline，确认字段、PBC、Jd、head/loss 无误且分域误差冲突确实存在。","2. 启用 escnmd_moe_backbone：初始 num_experts=4 或 8，moe_type=so2；数据少则减少专家或限制 layers_moe。","3. 绑定同源 elem_refs/normalizer/dataset_name；可选 DatasetSpecificMoEWrapper / dataset-specific head。","4. dry-run + 小 batch 冒烟后正式训练；监控整体与分 dataset 指标，以及 log_MOLE_stats 的专家系数均值/方差。","5. 输出 checkpoint 与分域误差对比，说明是否进入路由评估或合并部署。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_moe-routing-eval",title:"专家路由评估",summary:"统计专家利用率与分域误差，诊断塌缩或路由失效",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：多 dataset 验证集"},{id:"inputData",label:"MoE 模型与验证数据",placeholder:"例如：moe checkpoint + 各域 val.aselmdb",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：对比分域 MAE，检查专家系数方差与塌缩",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU"}],capabilitySummary:{current:"MoE 路由与分域精度评估",keyConfig:"expert_mixing_coefficients 均值/方差、分 dataset MAE、moe_dropout、专家数敏感性",resources:"GPU；按域拆 batch 避免 PBC 混合",deliverables:"分域误差表、专家利用率报告、是否保留 MoE 的结论"},requiredData:["训练好的 MoE checkpoint","按 dataset 可分组的验证集","训练时的 dataset_list 与路由字段约定"],expectedOutputs:"分域 MAE、专家系数分布、塌缩/过拟合诊断与调参建议",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MoE「专家路由评估」任务。","MoE 模型与验证数据：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MoE 评估约定执行：","1. 在验证集上计算整体与分 dataset 的 energy/force（及 stress）误差。","2. 收集 expert_mixing_coefficients，按 dataset 报告均值、方差与主导专家。","3. 诊断专家塌缩、某域路由失效或 composition embedding 元素覆盖问题。","4. 若分域无改善，建议退回 uma_escn_md 或调整 num_experts/layers_moe/dropout，而非继续盲目加专家。","5. 给出是否适合 merge 部署或继续动态 MoE 推理的明确结论。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"uma_escn_moe-deploy",title:"合并部署目标域",summary:"用代表性样本 merge_MOLE_model，得到固定上下文的普通 eSCNMDBackbone",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：固定 OMAT 生产域部署"},{id:"inputData",label:"MoE 模型与代表样本",placeholder:"例如：moe checkpoint + 目标域代表性 AtomicData（含 charge/spin/dataset）",multiline:!0},{id:"goal",label:"部署目标",placeholder:"例如：moe_type=so2 合并后与合并前同 batch 输出一致，再接弛豫",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 推理节点"}],capabilitySummary:{current:"merge_MOLE_model 固定域部署",keyConfig:"mole_type=so2、代表性 data 路由上下文、合并前后一致性、目标域 dataset",resources:"GPU；合并后推理成本接近普通 eSCNMD",deliverables:"合并后 eSCNMDBackbone、一致性报告、目标域弛豫/MD 入口"},requiredData:["moe_type=so2 的已训练 MoE 模型（merge 仅支持 so2）","含 atomic_numbers、batch、charge、spin、dataset 的代表性目标域样本","合并后验收用的同分布测试 batch"],expectedOutputs:"merged eSCNMDBackbone、一致性检查、部署脚本说明",promptTemplate:["请基于 {{system}} 体系完成 UMA-eSCN-MoE「合并部署目标域」任务。","MoE 模型与代表样本：{{inputData}}","部署目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MoE 部署路径执行：","1. 确认 mole_type/moe_type 为 so2；跨域动态泛化需求时保留动态 MoE，不要强行合并。","2. 选择目标域代表性 data_dict，调用 merged_model = model.merge_MOLE_model(data_dict)；merged_model.eval()。","3. 在同一 batch 上对比合并前后 energy/forces（及 stress）一致性。","4. 导出合并权重，并给出目标域 FAIRChemCalculator/Hydra 推理或短弛豫验证命令。","5. 说明合并模型仅适合该路由上下文，换 dataset 需重新合并或保留动态 MoE。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)}]},ee=[Y,Z,$],te={id:"fengwu",name:"FengWu",category:"onemodel",domain:"weather",tags:["气象","中期预报","多分支融合","全球大气"],blurb:"多分支中期天气预报：地表与高空 z/r/u/v/t 独立编码后跨变量融合，适合全球多气压层滚动预报。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/fengwu",intro:{title:"30 秒看懂模型",subtitle:"把 surface 与 z/r/u/v/t 六个变量族分开编码，在中分辨率三维变量轴上融合后再分支解码",suited:["整理六分支输入","中期多变量预报","评估跨变量技巧"],notFor:["缺六分支张量","分支网格不一致","只做单站点统计","天然单通道雷达短临"],prepare:["surface/z/r/u/v/t 张量","pressure_level 顺序","归一化均值/标准差","预训练权重"],obtain:["六分支预报场","滚动下一步输入","分变量 RMSE/ACC","NetCDF/Zarr 产品"],steps:["校验六分支与气压层","实例化 Fengwu","forward 单步冒烟","滚动重组输入","评估与落盘"]},install:{banner:"已为你创建 FengWu 环境配置任务",prompt:["请为 FengWu 创建一个可以直接跑通的环境配置任务。","模型定位：全球中期多分支天气预报；FengWuEncoder / FengWuFuser / FengWuDecoder，forward(surface, z, r, u, v, t)。","适合任务：整理六分支输入、中期多变量预报、评估跨变量技巧、配置权重与环境。","典型输入：surface=(B,4,H,W)；z/r/u/v/t=(B,pressure_level,H,W)；默认 img_size=(721,1440)、pressure_level=37、embed_dim=192、patch_size=(4,4)。","典型输出：六元组 (surface,z,r,u,v,t)；可滚动重组为下一步输入或写入 NetCDF/Zarr。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/fengwu：earth_env.sh、fake_data.py、train.py、inference.py、result.py、conf/config.yaml）与数据/权重路径。","3. 如果需要 ERA5、均值/标准差、静态场或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（例如 Fengwu(...) 对六个随机张量 forward 并打印 shape），并说明下一步可进入：整理六分支输入、中期多变量预报、评估跨变量技巧。"].join(`
-`)},scenarios:[{id:"fengwu-prepare-branches",title:"整理六分支输入",summary:"把再分析场拆成 surface 与 z/r/u/v/t，并对齐气压层与网格",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球 0.25° 大气、指定起报时次"},{id:"inputData",label:"输入数据",placeholder:"例如：ERA5 data_dir、global_means/stds、静态场、变量/气压层顺序",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成可直接 forward 的六分支张量与数据卡",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点整理 / 远程盘挂载"}],capabilitySummary:{current:"整理 FengWu 六分支输入",keyConfig:"surface 4 通道；z/r/u/v/t 通道=pressure_level；img_size、变量序、归一化统计量",resources:"登录节点或远程盘；大数据优先原地整理",deliverables:"六分支张量、字段/气压层检查报告、归一化约定"},requiredData:["ERA5/再分析或 fake_data 生成场","均值/标准差与静态文件（conf/config.yaml 的 data_dir）","地表变量序与高空压力层顺序说明"],expectedOutputs:"规范化 surface/z/r/u/v/t、shape 检查报告和可直接推理的数据入口",promptTemplate:["请基于 {{system}} 完成 FengWu「整理六分支输入」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FengWu 流程执行：","1. 从 ERA5/示例数据拆出 surface（默认 4 通道）与高空 z/r/u/v/t（每支通道数=pressure_level，默认 37）。","2. 校验六分支共享同一 Batch、Height、Width；默认网格 (721,1440)；缺测、归一化与网格对齐在调用前完成。","3. 记录变量顺序、压力层顺序、均值/标准差路径，形成短数据卡；禁止用全零张量冒充缺失高空族。","4. 对照 examples/earth/fengwu（fake_data.py / conf/config.yaml）给出下一步可直接用于推理的张量路径或构造命令。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fengwu-medium-range",title:"中期多变量预报",summary:"单步 forward 冒烟后按六分支重组输入做滚动中期预报",fields:a,capabilitySummary:{current:"运行 FengWu 中期多变量预报",keyConfig:"lead time、滚动步数、embed_dim/patch_size/window_size、设备与混合精度",resources:"推荐远程 GPU/DCU；全球默认分辨率显存占用高，可先单步冒烟",deliverables:"六分支预报场、滚动日志、中间检查点路径"},requiredData:["已对齐的 surface/z/r/u/v/t","与配置一致的预训练权重","目标 lead time 与是否滚动调用"],expectedOutputs:"六分支预报结果、滚动重组说明和推理日志",promptTemplate:["请基于 {{system}} 完成 FengWu「中期多变量预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FengWu 流程执行：","1. 实例化 Fengwu(img_size、pressure_level、embed_dim、patch_size、num_heads、window_size) 并加载权重。","2. 将六个张量放到同一设备，调用 forward(surface, z, r, u, v, t)，核对输出六元组 shape 与输入分支一致。","3. 先做单步 smoke test，再按变量族把输出重组为下一步输入，执行目标 lead time 的滚动预报。","4. 校验物理量范围与缺测，写出 result/output 或 NetCDF/Zarr，并说明 examples/earth/fengwu/inference.py 入口。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fengwu-evaluate",title:"评估跨变量技巧",summary:"分地表/高空变量族计算 RMSE、ACC 并定位弱技巧分支",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球大气、指定气压层子集"},{id:"inputData",label:"预报与真值",placeholder:"例如：result/output 六分支预报 + ERA5 真值 + 统计量",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断 z/t/u/v/r/surface 哪支需加强或重训",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 批量评估 / CPU 抽查"}],capabilitySummary:{current:"评估 FengWu 跨变量预报技巧",keyConfig:"分变量 RMSE/ACC、气压层分层、异常值后处理门槛",resources:"小样本 CPU 可抽查；全网格建议 GPU",deliverables:"分变量评估表、弱技巧分支结论、改进建议"},requiredData:["六分支预报场","对齐真值与归一化约定","评估时段与变量清单"],expectedOutputs:"分变量技巧表、异常场列表和是否继续滚动的结论",promptTemplate:["请基于 {{system}} 评估 FengWu 跨变量预报技巧。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请按分支计算 surface 与 z/r/u/v/t 的 RMSE/ACC（必要时按气压层分层），定位最大误差时次与弱技巧变量族；","检查反归一化、压力层顺序与训练约定是否一致，并对照 examples/earth/fengwu/result.py 给出可视化与改进建议。","若发现通道错位或网格不一致，请先暂停并给出最小修复方案。"].join(`
-`)},{id:"fengwu-setup",title:"配置权重与环境",summary:"核对 earth_env、示例入口、权重与六分支最小 forward",fields:a,capabilitySummary:{current:"配置 FengWu 运行环境",keyConfig:"earth_env.sh、conf/config.yaml、checkpoint、GPU/DCU",resources:"登录节点装依赖 + GPU 冒烟",deliverables:"环境说明、最小验证命令、下一步任务入口"},requiredData:["机器与队列权限","OneScience 与 examples/earth/fengwu 路径","权重来源"],expectedOutputs:"可复现的环境配置说明与最小验证输出",promptTemplate:["请为 {{system}} 配置 FengWu 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：","1. source examples/earth/earth_env.sh，确认 onescience.models.fengwu.Fengwu 可导入。","2. 核对 conf/config.yaml 的 data_dir、年份与权重路径；无真数据时可用 fake_data.py。","3. 用六个随机张量或 fake 数据做最小 forward，打印 [o.shape for o in y]。","4. 说明下一步可进入整理六分支输入、中期多变量预报或评估跨变量技巧。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},ae={id:"fourcastnet",name:"FourCastNet",category:"onemodel",domain:"weather",tags:["气象","AFNO","二维全球场","谱混合"],blurb:"AFNO 二维全球天气场模型：单时刻经纬场经 patch 频域混合后恢复，适合高吞吐全球场预报。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/fourcastnet",intro:{title:"30 秒看懂模型",subtitle:"单时刻 (B,C,H,W) 经 patch embedding 与 AFNO 频域混合，恢复同分辨率下一时刻全球场",suited:["准备二维天气场","AFNO 全球预报","评估谱预报误差"],notFor:["需要显式三维气压层结构","非规则网格","必须多历史时刻一次性输入","雷达分钟级短临"],prepare:["二维气象场 (B,C,H,W)","通道序与归一化","patch_size 可整除网格","AFNO 权重"],obtain:["同网格预测场","滚动下一步输入","RMSE/谱误差","可视化"],steps:["校验 img_size 与 patch","加载 FourCastNet","forward(x)","反归一化","滚动或评估"]},install:{banner:"已为你创建 FourCastNet 环境配置任务",prompt:["请为 FourCastNet 创建一个可以直接跑通的环境配置任务。","模型定位：二维 patch + AFNO 频域混合全球天气场模型；forward(x)，默认 in_chans=out_chans=19。","适合任务：准备二维天气场、AFNO 全球预报、评估谱预报误差、配置 AFNO 环境。","典型输入：(B,19,720,1440)；img_size=(720,1440)、patch_size=(8,8)、embed_dim=768、depth=12、num_blocks=8、sparsity_threshold=0.01。","典型输出：同分辨率 (B,out_chans,H,W) 预测场；可作滚动输入或指标评估。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/fourcastnet：earth_env.sh、fake_data.py、train.py、inference.py、result.py）与数据/权重路径。","3. 如果需要 ERA5 二维场、统计量或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（FourCastNet(...) 对随机 (1,19,720,1440) forward），并说明下一步可进入：准备二维天气场、AFNO 全球预报、评估谱预报误差。"].join(`
-`)},scenarios:[{id:"fourcastnet-prepare-grid",title:"准备二维天气场",summary:"整理经纬网格二维场、通道顺序，并保证 patch 可整除",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球 720×1440 二维大气场"},{id:"inputData",label:"输入数据",placeholder:"例如：ERA5 二维变量集、均值/标准差、通道清单",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 (B,19,H,W) 且 H/W 可被 patch 整除",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点 / 远程盘"}],capabilitySummary:{current:"准备 FourCastNet 二维输入场",keyConfig:"in_chans/out_chans、img_size 可被 patch_size 整除、通道序、归一化",resources:"登录节点或远程盘",deliverables:"规范化 (B,C,H,W)、通道卡、整除检查报告"},requiredData:["再分析/示例二维场","通道顺序与统计量","目标 img_size 与 patch_size"],expectedOutputs:"可直接 forward 的二维场与检查报告",promptTemplate:["请基于 {{system}} 完成 FourCastNet「准备二维天气场」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FourCastNet 流程执行：","1. 将数据整理为 (B,C,H,W)，默认 C=19、H=720、W=1440；完成标准化与缺测处理。","2. 校验 Height/Width 可被 patch_size（默认 8×8）整除，避免 token 与恢复网格不一致。","3. 固定通道顺序与训练一致，写出短数据卡；对照 examples/earth/fourcastnet/fake_data.py 与 conf。","4. 给出下一步可直接用于 AFNO 预报的张量或文件路径。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fourcastnet-afno-forecast",title:"AFNO 全球预报",summary:"加载权重后对二维场做 AFNO 推理，必要时滚动多步",fields:a,capabilitySummary:{current:"运行 FourCastNet/AFNO 全球预报",keyConfig:"patch_size、embed_dim、depth、num_blocks、sparsity_threshold、滚动步数",resources:"推荐远程 GPU；接口简单，适合批量二维场",deliverables:"预测场、推理日志、滚动输入说明"},requiredData:["(B,C,H,W) 输入场","匹配分辨率的权重与 pos_embed","预报步长"],expectedOutputs:"同网格预报场与推理日志",promptTemplate:["请基于 {{system}} 完成 FourCastNet「AFNO 全球预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FourCastNet 流程执行：","1. 实例化 FourCastNet(img_size、patch_size、in_chans、out_chans、embed_dim、depth、num_blocks、sparsity_threshold 等) 并加载权重。","2. 确认 embed_dim 可被 num_blocks 整除；位置编码长度等于 patch 数（默认 90×180=16200）。","3. 调用 forward(x) 得到 (B,out_chans,H,W)；先单步冒烟，再按目标步数滚动（输出作下一步输入）。","4. 反归一化后检查变量范围，保存结果并指向 examples/earth/fourcastnet/inference.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fourcastnet-evaluate",title:"评估谱预报误差",summary:"计算全球场 RMSE，并关注大尺度谱与稀疏阈值影响",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球二维场、关注风场/位势"},{id:"inputData",label:"预报与真值",placeholder:"例如：预报 .npy/NetCDF + 真值 + 归一化统计量",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断 AFNO 稀疏阈值是否损失小尺度信号",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估 / CPU 抽查"}],capabilitySummary:{current:"评估 FourCastNet 谱预报误差",keyConfig:"RMSE、频谱能量、sparsity_threshold 敏感性",resources:"全场建议 GPU；谱分析可子集抽样",deliverables:"误差表、谱对比图、阈值建议"},requiredData:["预报场与真值","通道映射","可选不同 sparsity 对照结果"],expectedOutputs:"RMSE/谱误差报告与改进建议",promptTemplate:["请基于 {{system}} 评估 FourCastNet 谱预报误差。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请计算分通道 RMSE，并对比大尺度谱能量；关注 sparsity_threshold / hard_thresholding_fraction 是否过度抑制小尺度天气信号。","核对反归一化与通道序，对照 examples/earth/fourcastnet/result.py 给出可视化与是否可继续滚动的结论。","若分辨率与 pos_embed 不匹配，请先暂停并给出插值或重训位置编码方案。"].join(`
-`)},{id:"fourcastnet-setup",title:"配置 AFNO 环境",summary:"检查依赖、权重与默认分辨率最小 forward",fields:a,capabilitySummary:{current:"配置 FourCastNet/AFNO 环境",keyConfig:"earth_env.sh、权重路径、img_size/patch_size",resources:"登录节点 + GPU 冒烟",deliverables:"环境说明与最小验证命令"},requiredData:["权限与 OneScience 路径","权重来源","目标分辨率"],expectedOutputs:"可复现环境说明与 shape 验证输出",promptTemplate:["请为 {{system}} 配置 FourCastNet 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：source examples/earth/earth_env.sh；导入 FourCastNet；用随机或 fake 数据做最小 forward 并打印 y.shape；","核对 examples/earth/fourcastnet 的 conf 与权重，说明下一步可进入准备二维天气场、AFNO 全球预报或评估谱预报误差。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},ie={id:"fuxi",name:"FuXi",category:"onemodel",domain:"weather",tags:["气象","时空预报","3D patch","目标时刻"],blurb:"多时间步三维 patch 预报：历史序列压入时空 embedding 后输出目标时刻全球场。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/fuxi",intro:{title:"30 秒看懂模型",subtitle:"固定长度历史 (B,C,T,H,W) 经 3D patch 压缩时间维，再经二维 U 形主干恢复单帧目标场",suited:["整理多时刻输入","目标时刻预报","评估时空预报"],notFor:["TimeSteps≠PatchTimeSteps","需要显式压力层三维大气主干","输出完整未来时间序列头","单帧且无历史窗"],prepare:["多时刻场 (B,C,T,H,W)","T=patch 时间长","in_chans/out_chans","FuXi 权重"],obtain:["单帧目标场 (B,out_chans,H,W)","反归一化产品","RMSE/ACC","可视化"],steps:["确认 T==patch_t","整理五维输入","forward(x)","插值恢复校验","评估落盘"]},install:{banner:"已为你创建 FuXi 环境配置任务",prompt:["请为 FuXi 创建一个可以直接跑通的环境配置任务。","模型定位：多时间步→单目标时刻的时空 patch 预报；FuxiEmbedding 压缩时间维后接 FuxiTransformer；forward(x)。","适合任务：整理多时刻输入、目标时刻预报、评估时空预报、配置 FuXi 环境。","典型输入：(B,70,2,721,1440)；img_size=(2,721,1440)、patch_size=(2,4,4)、embed_dim=1536、num_heads=8、window_size=7。","典型输出：(B,70,721,1440) 单帧场（patch 恢复后可能经 bilinear 插值）。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/fuxi：train_short/medium/long.py、inference.py、data_loader.py、fake_data.py）与数据/权重路径。","3. 如果需要多时刻 ERA5、统计量或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（Fuxi(...) 对 (1,70,2,721,1440) forward），并说明下一步可进入：整理多时刻输入、目标时刻预报、评估时空预报。"].join(`
-`)},scenarios:[{id:"fuxi-prepare-timesteps",title:"整理多时刻输入",summary:"把固定历史窗拼成 (B,C,T,H,W) 并满足 patch 时间约束",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球大气、短/中/长期窗配置"},{id:"inputData",label:"输入数据",placeholder:"例如：多时刻 ERA5、通道 70、T=2、均值/标准差",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 TimeSteps==PatchTimeSteps 的五维张量",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点 / 远程盘"}],capabilitySummary:{current:"整理 FuXi 多时刻输入",keyConfig:"TimeSteps、patch_size[0]、in_chans、网格 (721,1440)、归一化",resources:"登录节点或远程盘",deliverables:"五维张量、时间窗约定、检查报告"},requiredData:["连续历史时刻场","变量通道清单","目标 patch 时间长度"],expectedOutputs:"可 forward 的 (B,C,T,H,W) 与数据卡",promptTemplate:["请基于 {{system}} 完成 FuXi「整理多时刻输入」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FuXi 流程执行：","1. 将历史序列整理为 (B,C,T,H,W)，默认 C=70、T=2、H=721、W=1440。","2. 强制校验 TimeSteps == patch_size[0]（默认均为 2），否则初始化会失败。","3. 完成归一化、缺测填补与经纬对齐；记录短/中/长期任务对应的 examples/earth/fuxi 训练入口。","4. 给出下一步可直接用于目标时刻预报的张量路径。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fuxi-target-forecast",title:"目标时刻预报",summary:"对五维输入做 forward，得到单帧目标时刻全球场",fields:a,capabilitySummary:{current:"运行 FuXi 目标时刻预报",keyConfig:"embed_dim、window_size、out_chans、是否外层滚动多步",resources:"默认 embed_dim=1536 需远程 GPU；CPU 仅小尺寸调试",deliverables:"单帧预报场、推理日志、可选滚动计划"},requiredData:["五维输入张量","匹配配置的权重","目标 lead / 是否外层滚动"],expectedOutputs:"(B,out_chans,H,W) 预报与日志",promptTemplate:["请基于 {{system}} 完成 FuXi「目标时刻预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience FuXi 流程执行：","1. 实例化 Fuxi(img_size、patch_size、in_chans、out_chans、embed_dim、num_groups、num_heads、window_size) 并加载权重。","2. 调用 forward(x) 得到单帧 (B,out_chans,H,W)；注意当前实现不输出时间序列。","3. 检查 patch 恢复与 F.interpolate 是否掩盖尺寸不一致；对降水/风速等局地极值额外抽查。","4. 若需多步，在外层滚动重组时间窗；保存结果并指向 examples/earth/fuxi/inference.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"fuxi-evaluate",title:"评估时空预报",summary:"评估目标时刻场技巧，关注插值恢复对极值的影响",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球场、关注降水/风速"},{id:"inputData",label:"预报与真值",placeholder:"例如：目标时刻预报 + 真值 + 统计量",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断插值恢复是否平滑掉局地极值",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估"}],capabilitySummary:{current:"评估 FuXi 时空预报技巧",keyConfig:"RMSE/ACC、极值偏差、短/中/长期窗对比",resources:"全场建议 GPU",deliverables:"评估表、极值诊断、窗口选择建议"},requiredData:["预报与真值","时间窗配置","变量清单"],expectedOutputs:"技巧报告与是否更换短/中/长配置的结论",promptTemplate:["请基于 {{system}} 评估 FuXi 时空预报。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请计算目标时刻分通道 RMSE/ACC，抽查降水、风速等变量的局地极值是否被 bilinear 插值平滑；","对照 examples/earth/fuxi 的 short/medium/long 配置给出窗口与容量建议。","若发现 T 与 patch 时间不一致或通道错位，请先暂停并修复。"].join(`
-`)},{id:"fuxi-setup",title:"配置 FuXi 环境",summary:"核对依赖、短中长训练入口与五维最小 forward",fields:a,capabilitySummary:{current:"配置 FuXi 环境",keyConfig:"earth_env.sh、embed_dim、短/中/长脚本、权重",resources:"登录节点 + GPU 冒烟",deliverables:"环境说明与最小验证命令"},requiredData:["权限与路径","权重来源","目标时间窗配置"],expectedOutputs:"可复现环境说明与 shape 验证",promptTemplate:["请为 {{system}} 配置 FuXi 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：source examples/earth/earth_env.sh；确认 Fuxi 可导入；用默认五维随机张量最小 forward；","列出 train_short/medium/long.py 与 inference.py 入口，说明下一步任务。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},ne={id:"graphcast",name:"GraphCast",category:"onemodel",domain:"weather",tags:["气象","GNN","mesh","grid-to-mesh"],blurb:"网格-Mesh 图神经网络预报：经纬特征映射到多分辨率 icosahedral mesh 再解码回网格。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/graphcast",intro:{title:"30 秒看懂模型",subtitle:"经纬 grid 特征经 G2M 编码到 icosahedral mesh，消息传递后再 M2G 解码回全球网格",suited:["构建 grid-mesh","GNN 全球预报","评估 mesh 预报"],notFor:["必须大 batch>1 且无分区方案","无法接受构图/图依赖","只要二维 AFNO 单帧接口","雷达短临"],prepare:["grid_nfeat (1,C,H,W)","mesh_level/processor","input/output 特征维","图缓存与权重"],obtain:["(1,C_out,H,W) 网格预报","可选分区输出","技巧评估","自回归多步"],steps:["定特征维与 mesh_level","构图并实例化","forward(grid_nfeat)","prepare_output","评估/滚动"]},install:{banner:"已为你创建 GraphCast 环境配置任务",prompt:["请为 GraphCast 创建一个可以直接跑通的环境配置任务。","模型定位：经纬网格↔icosahedral mesh 的 GraphCastNet；forward(grid_nfeat)；可选 MessagePassing / GraphTransformer。","适合任务：构建 grid-mesh、GNN 全球预报、评估 mesh 预报、配置 GraphCast 环境。","典型输入：(1,237,721,1440)；mesh_level=6、multimesh=True、output_dim_grid_nodes=227、processor_layers=16、hidden_dim=512；非分布式 batch 必须为 1。","典型输出：聚合 (1,227,721,1440)；分布式下可为分区节点结果。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/graphcast：fake_data.py、compute_time_diff_std.py、get_data_json.py、train.py、inference.py、finetune.py）与数据/权重路径。","3. 如果需要 ERA5、time_diff_std、data json、图缓存或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（GraphCastNet(...) 对 (1,237,721,1440) forward），并说明下一步可进入：构建 grid-mesh、GNN 全球预报、评估 mesh 预报。"].join(`
-`)},scenarios:[{id:"graphcast-build-mesh",title:"构建 grid-mesh",summary:"按分辨率与 mesh_level 构图，并准备 grid 节点特征维",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球 0.25°、mesh_level=6"},{id:"inputData",label:"输入数据",placeholder:"例如：ERA5 网格特征、变量维 237→227、辅助 json/std",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成可缓存的 mesh/g2m/m2g 图与输入约定",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 构图 / 远程盘缓存图"}],capabilitySummary:{current:"构建 GraphCast grid-mesh 图",keyConfig:"input_res、mesh_level、multimesh、khop_neighbors、input/output_dim_grid_nodes",resources:"构图耗资源，建议远程盘缓存图对象",deliverables:"图缓存、特征维约定、辅助文件清单"},requiredData:["网格分辨率与变量集合","compute_time_diff_std / get_data_json 产出","目标 mesh_level 与是否 multimesh"],expectedOutputs:"图与特征约定、辅助文件路径、检查报告",promptTemplate:["请基于 {{system}} 完成 GraphCast「构建 grid-mesh」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphCast 流程执行：","1. 根据变量集合确定 input_dim_grid_nodes / output_dim_grid_nodes（默认 237 / 227）。","2. 选择 mesh_level、multimesh、khop_neighbors；运行 fake_data 或真实数据后执行 compute_time_diff_std.py 与 get_data_json.py。","3. 实例化时构建 mesh_graph / g2m_graph / m2g_graph，并规划图缓存，避免重复构图。","4. 明确非分布式 batch=1；若用分区，记录 partition_size 与 expect_partitioned_input 等互斥标志。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphcast-gnn-forecast",title:"GNN 全球预报",summary:"编码到 mesh、processor 传播后解码回经纬网格预报",fields:a,capabilitySummary:{current:"运行 GraphCast GNN 全球预报",keyConfig:"processor_type、processor_layers、hidden_dim、checkpoint、partition",resources:"推荐 GPU；显存不足时降 mesh_level/hidden_dim 或开 checkpoint/分区",deliverables:"网格预报、自回归日志、设备迁移说明"},requiredData:["grid_nfeat (1,C,H,W)","已构图模型与权重","lead / 自回归步数"],expectedOutputs:"(1,C_out,H,W) 预报与推理日志",promptTemplate:["请基于 {{system}} 完成 GraphCast「GNN 全球预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience GraphCast 流程执行：","1. 实例化 GraphCastNet（processor_type=MessagePassing 或 GraphTransformer；processor_layers≥3）并加载权重。","2. 用 model.to(...) 同步迁移模型、图与图特征；调用 forward(grid_nfeat)（内部 prepare_input→encoder→processor→decoder→prepare_output）。","3. 先单步冒烟确认输出 (1,output_dim_grid_nodes,H,W)，再按目标步数自回归；需要时设置 set_checkpoint_*。","4. 反归一化并检查范围，指向 examples/earth/graphcast/inference.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"graphcast-evaluate",title:"评估 mesh 预报",summary:"在经纬网格上评估技巧，并诊断图/特征维配置问题",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球中期、关注高层风场"},{id:"inputData",label:"预报与真值",placeholder:"例如：网格预报 + ERA5 真值 + 特征维映射",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断 mesh_level/processor 是否够用",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估"}],capabilitySummary:{current:"评估 GraphCast mesh 预报",keyConfig:"RMSE/ACC、按变量维切片、分区聚合一致性",resources:"全场 GPU；可先子集气压层/变量",deliverables:"技巧表、配置诊断、改进建议"},requiredData:["预报与真值","input/output 特征映射","评估时段"],expectedOutputs:"网格技巧报告与 mesh/processor 建议",promptTemplate:["请基于 {{system}} 评估 GraphCast mesh 预报。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请在经纬网格上按 output 特征维计算 RMSE/ACC；检查 batch、特征维与图分辨率是否与训练一致；","若分区模式，核对聚合输出是否完整。对照 examples/earth/graphcast/result.py 给出可视化与 mesh_level/processor 调整建议。"].join(`
-`)},{id:"graphcast-setup",title:"配置 GraphCast 环境",summary:"核对图依赖、辅助脚本与 batch=1 最小 forward",fields:a,capabilitySummary:{current:"配置 GraphCast 环境",keyConfig:"earth_env、图依赖、辅助脚本、checkpoint/分区开关",resources:"登录节点准备 + GPU 冒烟",deliverables:"环境说明与最小验证命令"},requiredData:["权限与路径","权重与辅助文件来源"],expectedOutputs:"可复现环境说明与 shape 验证",promptTemplate:["请为 {{system}} 配置 GraphCast 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：source examples/earth/earth_env.sh；跑通 fake_data 与辅助脚本入口；用 batch=1 张量最小 forward；","说明下一步可进入构建 grid-mesh、GNN 全球预报或评估 mesh 预报。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},re={id:"nowcastnet",name:"NowcastNet",category:"onemodel",domain:"weather",tags:["气象","短临降水","雷达","运动场"],blurb:"短临降水外推：雷达序列经运动场演化与生成式细节增强，输出分钟级未来降水。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/nowcastnet",intro:{title:"30 秒看懂模型",subtitle:"历史雷达帧经 Evolution_Network 预测运动/强度并 warp，再经生成编解码补充降水细节",suited:["准备雷达序列","短临降水外推","评估临近预报"],notFor:["全球中期多气压层预报","多通道未融合直接输入","只要气候尺度月季预测","无雷达/降水图像"],prepare:["雷达/降水序列","configs(total/input_length)","单通道图","mrms 权重"],obtain:["未来 pred_length 帧","可选 evo_result/motion","评分与产品阈值结果"],steps:["设 configs","整理 (B,T,H,W,C)","Net.forward","反缩放","业务阈值产品"]},install:{banner:"已为你创建 NowcastNet 环境配置任务",prompt:["请为 NowcastNet 创建一个可以直接跑通的环境配置任务。","模型定位：短临降水；Evolution_Network 运动/强度外推 + 生成式细节；Net.forward(all_frames)。","适合任务：准备雷达序列、短临降水外推、评估临近预报、配置短临环境。","典型输入：(B,29,256,256,1)；configs: total_length=29、input_length=9、pred_length=20、img_height/width=256、ngf=64、pretrained_model=./data/checkpoints/mrms_model.ckpt。","典型输出：未来 20 帧 gen_result；辅助 evolutionnet 可输出 evo_result / motion_。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/nowcastnet：requirements.txt、mrms_case_test.sh、mrms_case_train_evo/gen.sh、run.py、data/checkpoints）。","3. 如果需要 MRMS 数据、官方 data 目录或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（Net(cfg) 对随机序列 forward 并打印 shape），并说明下一步可进入：准备雷达序列、短临降水外推、评估临近预报。"].join(`
-`)},scenarios:[{id:"nowcastnet-prepare-radar",title:"准备雷达序列",summary:"整理单通道雷达/降水时序，满足长宽下采样与帧长配置",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：区域雷达拼图、对流个例"},{id:"inputData",label:"输入数据",placeholder:"例如：MRMS/雷达图序列、帧间隔、缩放约定",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 (B,29,H,W,1) 且 H/W 满足 /32,/8 约束",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点预处理"}],capabilitySummary:{current:"准备 NowcastNet 雷达序列",keyConfig:"total_length、input_length、img_height/width、单通道、缩放",resources:"本地或登录节点；大数据放 /data/dataset/mrms 一类路径",deliverables:"五维序列、configs 草案、尺寸检查报告"},requiredData:["雷达/降水图像序列","帧间隔与长度","目标高宽"],expectedOutputs:"可 forward 的 all_frames 与 configs 字段",promptTemplate:["请基于 {{system}} 完成 NowcastNet「准备雷达序列」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience NowcastNet 流程执行：","1. 将数据整理为 (B,total_length,H,W,C)，主模型只使用第一通道；默认示例 (1,29,256,256,1)。","2. 设置 input_length 与 total_length，使 pred_length=total_length-input_length；历史帧占前 input_length。","3. 校验高宽满足噪声投影 H/32、H/8 等约束；完成质量控制与缩放约定。","4. 对照 examples/earth/nowcastnet 数据布局给出下一步推理入口。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"nowcastnet-nowcast",title:"短临降水外推",summary:"演化外推后生成式增强，输出未来降水/回波序列",fields:a,capabilitySummary:{current:"运行 NowcastNet 短临外推",keyConfig:"ngf、device、pretrained_model、是否仅 evolution 基线、随机种子",resources:"推荐 GPU；CPU 仅小图调试；生成路径需固定种子以便复现",deliverables:"未来帧序列、可选运动场、产品阈值结果"},requiredData:["all_frames 序列","configs 与权重","是否要确定性 evo 基线"],expectedOutputs:"pred_length 帧预报与日志",promptTemplate:["请基于 {{system}} 完成 NowcastNet「短临降水外推」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience NowcastNet 流程执行：","1. 构造 configs（total_length、input_length、img_height/width、ngf、device、pretrained_model）并实例化 Net。","2. 确认 configs.device 与张量设备一致；调用 forward(all_frames) 得到 gen_result。","3. 需要稳健产品时，同时跑 evolutionnet 得到 evo_result/motion_ 作对照基线；固定随机种子保证可复现。","4. 反缩放（注意 pretrained_model 路径字符串可能影响 /128 逻辑），按业务阈值生成降水产品；参考 mrms_case_test.sh。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"nowcastnet-evaluate",title:"评估临近预报",summary:"评估短临技巧，对比演化基线与生成结果",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：强对流个例、区域雷达"},{id:"inputData",label:"预报与真值",placeholder:"例如：gen_result + evo_result + 观测序列",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断生成细节是否提升极端降水技巧",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估"}],capabilitySummary:{current:"评估 NowcastNet 临近预报",keyConfig:"逐帧 MAE/CSI、极端阈值、evo vs gen 对比",resources:"个例可 CPU；批量个例建议 GPU",deliverables:"逐帧技巧表、个例图、基线对比结论"},requiredData:["预报序列与观测","阈值定义","可选 evo 基线"],expectedOutputs:"短临评估报告与是否启用生成头的建议",promptTemplate:["请基于 {{system}} 评估 NowcastNet 临近预报。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请按未来帧计算误差/命中率（含极端降水阈值），对比 gen_result 与 evo_result；","检查单通道裁剪、缩放与随机性是否干扰结论，并给出产品策略建议。"].join(`
-`)},{id:"nowcastnet-setup",title:"配置短临环境",summary:"安装依赖、放置 data/checkpoints，并做最小序列 forward",fields:a,capabilitySummary:{current:"配置 NowcastNet 短临环境",keyConfig:"requirements.txt、data 目录、evo/gen 训练脚本、device",resources:"登录节点装依赖 + GPU 冒烟",deliverables:"环境说明与最小验证命令"},requiredData:["权限","MRMS/官方 data 与权重来源"],expectedOutputs:"可复现环境说明与 shape 验证",promptTemplate:["请为 {{system}} 配置 NowcastNet 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：pip install -r examples/earth/nowcastnet/requirements.txt；放置 data/checkpoints；","用 SimpleNamespace configs 做最小 Net.forward；说明 mrms_case_test/train 脚本下一步入口。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},oe={id:"pangu",name:"Pangu",category:"onemodel",domain:"weather",tags:["气象","三维大气","地表高空","Swin"],blurb:"三维大气统一主干：地表加静态掩码与高空变量拼接后做 Swin 融合，输出地表与高空场。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/pangu_weather",intro:{title:"30 秒看懂模型",subtitle:"单张量拼接地表/静态掩码/高空后，surface 与 upper-air patch 沿气压维合并进统一 3D trunk",suited:["整理三维大气输入","三维大气预报","评估大气技巧"],notFor:["天然六分支不愿拼通道","缺静态掩码","只要二维 AFNO","雷达短临"],prepare:["拼接张量 (B,72,H,W)","4+3+5×13 通道布局","13 气压层","Pangu 权重"],obtain:["地表 (B,4,H,W)","高空 (B,5,13,H,W)","滚动下一步输入","技巧评估"],steps:["按通道拼接输入","校验 recovery 网格","forward(x)","重组静态掩码","评估/滚动"]},install:{banner:"已为你创建 Pangu 环境配置任务",prompt:["请为 Pangu 创建一个可以直接跑通的环境配置任务。","模型定位：Pangu-Weather 风格三维全球预报；地表与高空 patch 沿气压维拼接后统一 3D Swin/窗口融合；forward(x)->(ys,yu)。","适合任务：整理三维大气输入、三维大气预报、评估大气技巧、配置 Pangu 环境。","典型输入：(B,72,721,1440)，72=4 地表预报 +3 静态掩码 +5×13 高空；patch_size=(2,4,4)、embed_dim=192、window_size=(2,6,12)。","典型输出：output_surface=(B,4,H,W)，output_upper_air=(B,5,13,H,W)；静态掩码不在输出中恢复。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/pangu_weather：earth_env.sh、fake_data.py、train.py、inference.py、result.py、scnet_pangu.py）。","3. 如果需要 ERA5、静态掩码、统计量或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（Pangu(...) 对 (1,72,721,1440) forward 打印 ys/yu shape），并说明下一步可进入：整理三维大气输入、三维大气预报、评估大气技巧。"].join(`
-`)},scenarios:[{id:"pangu-prepare-3d",title:"整理三维大气输入",summary:"按 4+3+5×13 通道布局拼接地表、静态掩码与高空场",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球 721×1440、13 气压层"},{id:"inputData",label:"输入数据",placeholder:"例如：ERA5 地表/高空、3 通道静态掩码、均值/标准差",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 (B,72,H,W) 并写清通道切片约定",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点 / 远程盘"}],capabilitySummary:{current:"整理 Pangu 三维大气输入",keyConfig:"通道切片 4+3+65、气压层 13、img_size、归一化",resources:"登录节点或远程盘",deliverables:"拼接张量、通道卡、静态掩码持续供给说明"},requiredData:["地表与高空再分析","静态掩码 3 通道","变量/气压层顺序"],expectedOutputs:"(B,72,H,W) 与通道布局数据卡",promptTemplate:["请基于 {{system}} 完成 Pangu「整理三维大气输入」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Pangu 流程执行：","1. 按约定拼接：前 4 通道地表预报变量、接着 3 通道静态掩码、剩余 65 通道为 5 类×13 层高空展平。","2. 校验默认网格 (721,1440) 与 recovery 绑定；完成归一化与缺测处理。","3. 写出通道切片卡：滚动时静态掩码必须从外部持续提供，不会出现在模型输出中。","4. 对照 examples/earth/pangu_weather/fake_data.py 与 conf 给出可推理路径。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"pangu-3d-forecast",title:"三维大气预报",summary:"统一 3D trunk 推理，得到地表与高空预报并支持滚动",fields:a,capabilitySummary:{current:"运行 Pangu 三维大气预报",keyConfig:"patch_size、window_size、embed_dim、lead、滚动重组",resources:"推荐远程 GPU；三维窗口融合显存高",deliverables:"地表/高空预报、滚动输入、推理日志"},requiredData:["(B,72,H,W) 输入","权重","目标 lead / 滚动步数"],expectedOutputs:"ys/yu 预报场与日志",promptTemplate:["请基于 {{system}} 完成 Pangu「三维大气预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Pangu 流程执行：","1. 实例化 Pangu(img_size、patch_size、embed_dim、num_heads、window_size) 并加载权重。","2. 调用 forward(x) 得到 ys=(B,4,H,W)、yu=(B,5,13,H,W)；先单步冒烟。","3. 滚动时把预报地表/高空与外部静态掩码重新拼成下一步 (B,72,H,W)。","4. 校验物理范围并保存；指向 examples/earth/pangu_weather/inference.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"pangu-evaluate",title:"评估大气技巧",summary:"分地表与高空气压层评估 RMSE/ACC",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球大气、指定气压层"},{id:"inputData",label:"预报与真值",placeholder:"例如：ys/yu + ERA5 真值 + 统计量",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断地表与高层风场技巧是否达标",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估"}],capabilitySummary:{current:"评估 Pangu 大气预报技巧",keyConfig:"分地表/高空 RMSE/ACC、气压层分层、通道错位检查",resources:"全场建议 GPU",deliverables:"分层评估表、错位诊断、改进建议"},requiredData:["ys/yu 预报","真值与归一化","气压层清单"],expectedOutputs:"地表/高空技巧报告与结论",promptTemplate:["请基于 {{system}} 评估 Pangu 大气预报技巧。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请分别评估地表 4 通道与高空 (5,13) 场；重点检查通道顺序错误导致的“shape 正确但语义错”问题。","对照 examples/earth/pangu_weather/result.py 给出可视化与是否继续滚动的结论。"].join(`
-`)},{id:"pangu-setup",title:"配置 Pangu 环境",summary:"核对 earth_env、权重与 (1,72,H,W) 最小 forward",fields:a,capabilitySummary:{current:"配置 Pangu 环境",keyConfig:"earth_env.sh、recovery 网格、权重路径",resources:"登录节点 + GPU 冒烟",deliverables:"环境说明与最小验证命令"},requiredData:["权限与路径","权重与静态掩码来源"],expectedOutputs:"可复现环境说明与 ys/yu shape 验证",promptTemplate:["请为 {{system}} 配置 Pangu 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：source examples/earth/earth_env.sh；导入 Pangu；对 (1,72,721,1440) 最小 forward 打印 ys/yu.shape；","说明下一步可进入整理三维大气输入、三维大气预报或评估大气技巧。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},se={id:"xihe",name:"Xihe",category:"onemodel",domain:"weather",tags:["气象","海洋预报","涡旋分辨","海陆掩膜"],blurb:"海洋涡旋分辨预报：规则海洋场配合海陆掩膜多尺度融合，适合高分辨率全球海洋预测。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/earth/xihe",intro:{title:"30 秒看懂模型",subtitle:"高分辨率海洋二维场经 patch embedding，在带海陆掩膜的 XiheFuser 多尺度 U 形主干中预报",suited:["准备海洋输入场","海洋涡旋预报","评估海洋技巧"],notFor:["无海陆掩膜的生产推理","大气六分支中期为主","雷达短临降水","非规则非结构化网格"],prepare:["海洋场 (B,in_chans,H,W)","land_mask.npy","config 尺寸/通道","Xihe 权重"],obtain:["海洋预报场 (B,out_chans,H,W)","仅海洋区域产品","技巧评估","可视化"],steps:["校验 mask 与网格","构造 config","forward(x)","反归一化与掩膜后处理","评估"]},install:{banner:"已为你创建 Xihe 环境配置任务",prompt:["请为 Xihe 创建一个可以直接跑通的环境配置任务。","模型定位：全球高分辨率海洋涡旋分辨预报；XiheEmbedding + 带 mask 的 XiheFuser U 形主干；forward(x)。","适合任务：准备海洋输入场、海洋涡旋预报、评估海洋技巧、配置 Xihe 环境。","典型输入：(B,96,2041,4320)；config: img_size=(2041,4320)、patch_size=(6,12)、in_chans=96、out_chans=94、embed_dim=192、mask=land_mask.npy。","典型输出：(B,94,H,W) 海洋变量场；应用海陆掩膜做后处理。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience demo（examples/earth/xihe：earth_env.sh、fake_data.py、train.py、inference.py、result.py、conf/config.yaml 中 data_dir/stats_dir/mask）。","3. 如果需要海洋年度 h5、global_means/stds、land_mask.npy 或 checkpoint，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证（Xihe(config=...) 对随机海洋张量 forward），并说明下一步可进入：准备海洋输入场、海洋涡旋预报、评估海洋技巧。"].join(`
-`)},scenarios:[{id:"xihe-prepare-ocean",title:"准备海洋输入场",summary:"整理海洋变量、统计量与海陆掩膜，保证 mask 与网格一致",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球涡旋分辨海洋、区域海盆"},{id:"inputData",label:"输入数据",placeholder:"例如：data/{year}.h5、global_means/stds、land_mask.npy",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：生成 (B,96,H,W) 与可读 mask，写入 config",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：登录节点 / 远程盘"}],capabilitySummary:{current:"准备 Xihe 海洋输入",keyConfig:"in_chans/out_chans、img_size、patch_size、mask 路径、归一化",resources:"高分辨率场优先远程盘；mask 必须可读",deliverables:"海洋张量、mask 校验、config 草案"},requiredData:["海洋变量场或 fake_data","land_mask.npy","均值/标准差"],expectedOutputs:"可 forward 的海洋输入与 config 字段",promptTemplate:["请基于 {{system}} 完成 Xihe「准备海洋输入场」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Xihe 流程执行：","1. 整理海洋变量为 (B,in_chans,H,W)，默认 in_chans=96、网格可至 (2041,4320)。","2. 校验 mask 文件存在且二维 shape 与网格对应；海洋区为正值或 >0.5。","3. 配置 conf：data_dir、stats_dir、mask、年份；完成归一化。","4. 对照 examples/earth/xihe/fake_data.py 给出下一步预报入口。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"xihe-eddy-forecast",title:"海洋涡旋预报",summary:"带海陆掩膜多尺度融合推理，输出高分辨率海洋场",fields:a,capabilitySummary:{current:"运行 Xihe 海洋涡旋预报",keyConfig:"embed_dim、window_size、patch_size、多尺度 mask 降采样、显存预算",resources:"默认高分辨率显存压力大，必须远程 GPU；可降分辨率冒烟",deliverables:"海洋预报场、掩膜后处理产品、推理日志"},requiredData:["海洋输入张量","config.mask 与权重","目标时效"],expectedOutputs:"(B,out_chans,H,W) 与日志",promptTemplate:["请基于 {{system}} 完成 Xihe「海洋涡旋预报」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience Xihe 流程执行：","1. 用 SimpleNamespace/配置对象提供 img_size、patch_size、mask、in_chans、out_chans、embed_dim 等，实例化 Xihe。","2. 调用 forward(x)；内部 change_mask 会为各尺度生成 TensorWithMask 供 XiheFuser 使用。","3. 先小分辨率或 fake 数据冒烟，再上正式高分辨率；注意奇偶 patch 网格边界。","4. 反归一化并对陆地区域掩膜后处理；指向 examples/earth/xihe/inference.py。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"xihe-evaluate",title:"评估海洋技巧",summary:"在有效海洋区域评估变量技巧与涡旋结构",fields:[{id:"system",label:"预报对象 / 区域",placeholder:"例如：全球海洋、关注 SST/SSH/流速"},{id:"inputData",label:"预报与真值",placeholder:"例如：预报场 + 真值 + land_mask + 统计量",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断中尺度涡旋与关键变量技巧",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：GPU 评估"}],capabilitySummary:{current:"评估 Xihe 海洋预报技巧",keyConfig:"仅海洋区 RMSE/ACC、分变量、涡旋结构诊断",resources:"全场 GPU；可先区域子集",deliverables:"分变量技巧表、涡旋个例图、结论"},requiredData:["预报与真值","海陆掩膜","变量清单"],expectedOutputs:"海洋区评估报告与改进建议",promptTemplate:["请基于 {{system}} 评估 Xihe 海洋预报技巧。","预报与真值：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请仅在 mask 海洋区计算分变量误差，抽查海表高度/温度/流速等涡旋相关结构；","对照 examples/earth/xihe/result.py 给出可视化，并判断是否需调整 patch/显存/分辨率配置。"].join(`
-`)},{id:"xihe-setup",title:"配置 Xihe 环境",summary:"核对 earth_env、mask 路径与高分辨率最小 forward",fields:a,capabilitySummary:{current:"配置 Xihe 环境",keyConfig:"earth_env.sh、mask.npy、stats、显存策略",resources:"登录节点准备 + GPU 冒烟（可降分辨率）",deliverables:"环境说明与最小验证命令"},requiredData:["权限与路径","mask/权重来源"],expectedOutputs:"可复现环境说明与 shape 验证",promptTemplate:["请为 {{system}} 配置 Xihe 运行环境。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请完成：source examples/earth/earth_env.sh；确认 conf 中 mask/stats 可读；","用 fake_data 或降分辨率配置做最小 forward；说明下一步可进入准备海洋输入场、海洋涡旋预报或评估海洋技巧。","如需安装依赖或提交远程命令，请先向我请求批准。"].join(`
-`)}]},le=[te,ae,ie,ne,re,oe,se],de=[...ee,...le,...K,...O],i=[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：CuO、水体系、界面结构"},{id:"inputData",label:"输入数据",placeholder:"例如：AIMD 轨迹、POSCAR、in.lammps",multiline:!0},{id:"goal",label:"任务目标",placeholder:"希望得到什么结果",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / HPC 队列"}],pe=[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：CuO、水体系、界面结构"},{id:"inputData",label:"输入数据",placeholder:"例如：aimd_dataset.extxyz、XDATCAR/OUTCAR 转换结果",multiline:!0},{id:"goal",label:"任务目标",placeholder:"例如：训练可用于材料模拟的 MACE 势函数",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：远程 GPU / 单卡冒烟 / HPC 队列"}],me={id:"mace",name:"MACE",category:"onemodel",domain:"materials",tags:["材料化学","mlff","equivariant"],blurb:"高阶等变消息传递机器学习势函数，用 AIMD/DFT 轨迹训练后可接入大规模分子动力学。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/mace",intro:{title:"30 秒看懂模型",subtitle:"用 AIMD/DFT 轨迹数据训练原子间势函数，用于更快地做大规模分子动力学模拟",suited:["训练力场","结构弛豫","分子动力学","能量预测"],notFor:["没有训练数据","体系差异过大","只想做电子结构"],prepare:["AIMD 轨迹","DFT 结构","能量和力数据"],obtain:["训练好的势函数","预测结果","MD 轨迹"],steps:["整理 AIMD/DFT 数据","检查字段与单位","训练或微调 MACE","评估误差","接入 LAMMPS/ASE"]},install:{banner:"已为你创建 MACE 环境配置任务",prompt:["请为 MACE 创建一个可以直接跑通的环境配置任务。","模型定位：E(3)-等变高阶消息传递机器学习势函数。","适合任务：整理 AIMD 训练数据、训练/微调势函数、评估能量/力误差、ASE/LAMMPS 部署与 MD。","典型输入：extxyz/xyz 构型、energy_key/forces_key、可选 stress/virial、元素表与 E0s。","典型输出：checkpoint、误差表、导出模型、LAMMPS/ASE 验证结果。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU/DCU、远程环境和项目目录权限。","2. 优先核对 OneScience MACE demo（examples/matchem/mace/demo）与数据路径是否可用。","3. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","4. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","5. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：整理 AIMD 训练数据、训练机器学习势、评估势函数精度、准备 LAMMPS 分子动力学。"].join(`
-`)},scenarios:[{id:"mace-train",title:"训练机器学习势",summary:"用 AIMD 或 DFT 轨迹训练 MACE 势函数",fields:pe,capabilitySummary:{current:"训练 MACE 机器学习势",keyConfig:"r_max/cutoff、batch size、epoch、energy_weight/forces_weight、E0s、可选 foundation_model",resources:"建议远程 GPU/DCU；小样本可先单卡冒烟，多卡用 demo YAML + dry-run",deliverables:"checkpoint、训练曲线、能量/力 MAE 和部署说明"},requiredData:["含能量/力标签的 .xyz / .extxyz（train，推荐 valid/test）","energy_key / forces_key（示例 TolEnergy / force），单位与 E0s 约定","元素表；微调时还需 foundation checkpoint 与兼容 r_max"],expectedOutputs:"checkpoint、训练曲线、能量/力 MAE 和部署说明",promptTemplate:["请基于 {{system}} 体系完成 MACE「训练机器学习势」任务。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","","请按 OneScience MACE 流程执行：","1. 检查 extxyz/xyz 字段是否覆盖结构、能量、力、晶胞；确认 energy_key、forces_key、可选 stress_key/virials_key。","2. 明确元素表、能量/力单位与 E0s（isolated / average / foundation / 显式字典）。","3. 选择路线：数据少或接近预训练覆盖则 foundation_model 微调；数据充足或体系差异大则从头训练。","4. 先用 demo YAML/`run.sh --dry-run` 核对命令与路径，再做短训练 smoke test，确认 forward、loss、checkpoint 正常。","5. 正式训练后输出能量/力 MAE、异常帧、checkpoint 路径，并说明下一步可做评估或 LAMMPS/ASE 部署。","如需安装依赖、加载模块或提交远程命令，请先向我请求批准。"].join(`
-`)},{id:"mace-evaluate",title:"评估势函数精度",summary:"比较 MACE 预测与 DFT 标签的误差和稳定性",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：CuO、水体系、界面结构"},{id:"inputData",label:"模型与测试数据",placeholder:"例如：checkpoints/*.model + water_test.xyz",multiline:!0},{id:"goal",label:"评估目标",placeholder:"例如：判断能否替代 DFT 采样进入 MD",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：本地 CPU / 远程 GPU"}],capabilitySummary:{current:"评估现有 MACE 势函数",keyConfig:"能量/力 MAE、RMSE、PerAtomMAE、异常帧筛选、可选 stress/virial 检查",resources:"小规模 CPU 即可；大数据集或含应力评估建议 GPU",deliverables:"误差表、异常结构列表和是否可用于 MD 的结论"},requiredData:["待评估 checkpoint / .model","独立测试集 extxyz（字段与训练一致）","训练时使用的 energy_key、forces_key、元素表与单位约定"],expectedOutputs:"误差表、异常结构列表和是否可用于 MD 的结论",promptTemplate:["请基于 {{system}} 体系评估 MACE 势函数精度。","模型与测试数据：{{inputData}}","评估目标：{{goal}}","运行环境：{{environment}}","","请在独立测试集上计算能量和力的 MAE / RMSE（必要时按 PerAtomMAE），定位最大误差帧与异常近邻构型，","检查单位、E0s 与字段映射是否与训练一致，并给出能否进入 ASE/LAMMPS MD 或结构弛豫的明确结论。","如果发现字段不匹配或单位错配，请先暂停并给出最小修复方案。"].join(`
-`)},{id:"mace-lammps",title:"准备 LAMMPS 分子动力学",summary:"把训练好的 MACE 势函数用于 LAMMPS",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：CuO、水体系、界面结构"},{id:"inputData",label:"模型与初始结构",placeholder:"例如：trained.model + data.lammps / POSCAR 转换结果",multiline:!0},{id:"goal",label:"MD 目标",placeholder:"例如：300K NVT 短时程 smoke test 后拉长采样",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：HPC GPU/CPU 队列"}],capabilitySummary:{current:"把 MACE 模型接入 LAMMPS",keyConfig:"pair_style mace、温度压力、步数、dump/restart、导出模型与元素顺序",resources:"GPU 或 CPU 队列；先做约 10 ps smoke test 再拉长",deliverables:"LAMMPS 输入文件、作业脚本、轨迹与能量输出"},requiredData:["已训练/导出的 MACE 模型","初始结构（LAMMPS data 或可转换结构）","温度/压力/系综与输出需求（dump、thermo、restart）"],expectedOutputs:"LAMMPS 输入文件、作业脚本、轨迹与能量输出",promptTemplate:["请基于 {{system}} 体系，把 MACE 势函数接入 LAMMPS 做分子动力学。","模型与初始结构：{{inputData}}","MD 目标：{{goal}}","运行环境：{{environment}}","","请先完成短时程 smoke test：确认模型加载、元素顺序、单位、周期边界、温度/能量、dump 和 restart 正常；","再与 ASE/原训练模型在同一构型上做一致性抽查，最后给出更长时程 MD 计划与后处理入口。","如需提交远程作业或安装 LAMMPS/MACE 插件，请先向我请求批准。"].join(`
-`)},{id:"mace-prepare-data",title:"整理 AIMD 训练数据",summary:"从 VASP 或 AIMD 输出生成训练数据集",fields:[{id:"system",label:"研究对象 / 材料体系",placeholder:"例如：CuO、水体系、界面结构"},{id:"inputData",label:"原始轨迹 / 计算输出",placeholder:"例如：XDATCAR、OUTCAR、vasprun.xml、已有 xyz 片段",multiline:!0},{id:"goal",label:"整理目标",placeholder:"例如：生成带 TolEnergy/force 的 train/valid/test.extxyz",multiline:!0},{id:"environment",label:"运行环境",placeholder:"例如：本地或登录节点"}],capabilitySummary:{current:"整理 MACE 训练所需 AIMD/DFT 数据",keyConfig:"energy_key/forces_key、单位、E0s、元素表、train/valid/test 划分",resources:"本地或登录节点即可；大数据优先远程盘原地整理",deliverables:"extxyz 数据集、数据卡、字段/单位检查报告"},requiredData:["VASP/AIMD 输出或中间轨迹（XDATCAR/OUTCAR/vasprun 等）","DFT 设置与能量/力单位说明","目标字段名（如 TolEnergy、force）和划分比例"],expectedOutputs:"extxyz 训练集、验证/测试划分、字段检查报告和数据卡",promptTemplate:["请为 {{system}} 体系整理可用于 MACE 训练的 AIMD/DFT 数据集。","原始轨迹 / 计算输出：{{inputData}}","整理目标：{{goal}}","运行环境：{{environment}}","","请完成：","1. 从 VASP/AIMD 输出提取结构、能量、力、晶胞，必要时做清洗与异常帧过滤。","2. 写出 .xyz/.extxyz，明确 energy_key、forces_key，以及可选 stress_key/virials_key。","3. 核对单位、元素表、周期边界与参考能策略，形成简短数据卡。","4. 按约定划分 train/valid/test，并给出下一步可直接用于 MACE 训练的命令入口。"].join(`
-`)}]},ce={id:"deepmd",name:"DeePMD",category:"onemodel",domain:"materials",tags:["材料化学","mlff","deepmd"],blurb:"深度学习势能函数，用 DFT 标注的能量、力和晶胞数据训练后可接入 LAMMPS。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/dp",intro:{title:"30 秒看懂模型",subtitle:"用 DFT 标注的能量、力和晶胞数据训练 Deep Potential 力场，并可接入 LAMMPS 做大规模分子动力学",suited:["Deep Potential 训练","AIMD 数据学习","LAMMPS 推理","主动学习闭环"],notFor:["没有能量/力标签","数据格式无法转为 deepmd/npy","只需要单点电子结构结果"],prepare:["VASP/CP2K 输出","deepmd/npy 数据目录","结构、能量、力和 virial"],obtain:["frozen_model.pb / saved_model","训练曲线","测试误差","LAMMPS pair_style 配置"],steps:["转换训练数据","配置 input.json","训练 Deep Potential","冻结模型","接入 LAMMPS 验证"]},install:{banner:"已为你创建 DeePMD 环境配置任务",prompt:["请为 DeePMD 创建一个可以直接跑通的环境配置任务。","模型定位：深度势能分子动力学模型。","适合任务：Deep Potential 训练、AIMD 数据学习、LAMMPS 推理、主动学习闭环。","典型输入：VASP/CP2K 输出 / deepmd/npy 数据目录 / 结构、能量、力和 virial。","典型输出：frozen_model.pb / 训练曲线 / 测试误差 / LAMMPS pair_style 配置。","要求：","1. 先检查当前项目、Python/框架版本、CPU/GPU、远程环境和项目目录权限。","2. 如果需要模型权重、训练数据或插件，请列出来源、版本和下载/挂载方式。","3. 如果需要安装依赖、加载模块或执行远程命令，请先向我请求批准。","4. 安装/配置完成后，请给出最小验证命令，并说明下一步可以进入：转换 DeepMD 数据、训练 Deep Potential、部署到 LAMMPS。"].join(`
-`)},scenarios:[{id:"deepmd-convert",title:"转换 DeepMD 数据",summary:"把 VASP/CP2K/XYZ 数据转成 deepmd/npy 训练目录",fields:i,capabilitySummary:{current:"准备 DeePMD 训练数据",keyConfig:"coord、box、energy、force，可选 virial",resources:"本地或远程均可；大数据优先远程盘引用",deliverables:"可被 dp train 直接读取的 deepmd/npy 数据目录"},expectedOutputs:"deepmd/npy 目录、type_map、数据检查报告",promptTemplate:["请把 {{system}} 体系的数据转换成 DeePMD 训练所需的 deepmd/npy 结构。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请检查能量、力和晶胞信息是否齐全，并保证 type.raw / type_map.raw 一致。"].join(`
-`)},{id:"deepmd-train",title:"训练 Deep Potential",summary:"基于 input.json 训练 DeepMD 力场",fields:i,capabilitySummary:{current:"训练 DeePMD 势函数",keyConfig:"descriptor、fitting_net、learning_rate、batch_size、loss 权重",resources:"通常建议远程 GPU；小样本可先用 CPU/GPU 单卡验证",deliverables:"checkpoint、训练曲线、lcurve.out 和模型可用性判断"},expectedOutputs:"训练日志、模型 checkpoint、能量/力误差和验证建议",promptTemplate:["请基于 {{system}} 体系训练 DeePMD Deep Potential。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请生成 input.json、启动训练，并输出误差曲线和模型文件。"].join(`
-`)},{id:"deepmd-lammps",title:"部署到 LAMMPS",summary:"冻结模型并生成 LAMMPS 调用脚本",fields:i,capabilitySummary:{current:"把 DeePMD 模型接入 LAMMPS",keyConfig:"pair_style deepmd、温度压力、步数、dump/restart",resources:"GPU 或 CPU 队列；先做短时程 smoke test",deliverables:"frozen_model、LAMMPS 输入文件、作业脚本与轨迹输出"},expectedOutputs:"frozen_model.pb、LAMMPS 输入、作业脚本、轨迹与能量输出",promptTemplate:["请基于 {{system}} 体系，把 DeePMD 势函数冻结并接入 LAMMPS。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请完成模型冻结、生成 LAMMPS 调用脚本，并先做短时程 smoke test 确认 pair_style deepmd、能量和 dump 正常。"].join(`
-`)}]},ue={id:"lammps",name:"LAMMPS",category:"hpc",domain:"materials",tags:["HPC","分子动力学","mlff"],blurb:"大规模分子动力学软件，适合经典力场和机器学习势驱动的长时间、大体系模拟。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/tools/lmp",intro:{title:"30 秒看懂 HPC 软件",subtitle:"用于经典力场和机器学习势驱动的分子动力学模拟，适合大体系和长时间尺度计算",suited:["分子动力学","MACE/DeePMD 力场调用","热力学性质","RDF/MSD 后处理"],notFor:["没有可用力场","初始结构不稳定","边界条件和单位体系不明确"],prepare:["data 文件","in.* 输入脚本","力场文件","温度/压力/步数"],obtain:["dump 轨迹","log.lammps","thermo 数据","RDF/MSD 图表"],steps:["选择力场","生成 LAMMPS 输入","检查单位和边界","提交 MD 任务","分析轨迹和热力学量"]},install:{banner:"已为你创建 LAMMPS 软件环境配置任务",prompt:["请检查当前项目和远程 HPC 环境是否可以运行 LAMMPS，并完成软件环境配置方案。","软件定位：大规模分子动力学软件。","版本：lammps-mace","MACE 接口需要 conda 环境。","conda 环境：curl -f -C - -o lammps-mace.tar.gz https://zzefile.scnet.cn:65011/efile/s/d/emhhbmdobDM=/8b21039fd3c245e6","二进制文件链接：curl -f -C - -o lammps-mace-dtk2604.tar.gz https://zzefile.scnet.cn:65011/efile/s/d/emhhbmdobDM=/861d7dea382b0a71","典型输入：data 文件 / in.* 输入脚本 / 力场文件 / 温度压力步数。","典型输出：dump 轨迹 / log.lammps / thermo 数据 / RDF/MSD 图表。","要求：","1. 先检查 module、conda/container、MPI/GPU、许可证/访问权限和项目目录权限。","2. 说明 LAMMPS 适用的运行方式、推荐版本、依赖环境和验证命令。","3. 如果需要安装依赖、加载模块或提交远程命令，请先向我请求批准。","4. 安装/配置完成后，请给出最小验证用例，并说明后续可以做：MACE 力场 MD、经典力场 MD、轨迹后处理。"].join(`
-`)},scenarios:[{id:"lammps-mace-md",title:"MACE 力场 MD",summary:"把训练好的 MACE 模型接入 LAMMPS 做分子动力学",fields:i,capabilitySummary:{current:"LAMMPS + MACE 分子动力学",keyConfig:"pair_style mace、ensemble、温度压力、dump 间隔",resources:"HPC CPU/GPU 队列；先短时程验证再拉长",deliverables:"in.mace_md、提交脚本、dump 轨迹和 thermo 日志"},expectedOutputs:"in.mace_md、run_lammps.slurm、dump 轨迹、log 文件和后处理入口",promptTemplate:["请为 {{system}} 体系配置 LAMMPS MACE 力场分子动力学。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请生成输入脚本、检查单位与边界，并给出提交与后处理计划。"].join(`
-`)},{id:"lammps-classical-md",title:"经典力场 MD",summary:"使用 EAM、ReaxFF、LJ 等经典力场配置模拟",fields:i,capabilitySummary:{current:"LAMMPS 经典力场分子动力学",keyConfig:"pair_style、ensemble、温度压力、单位与边界条件",resources:"HPC CPU/GPU 队列；先短时程验证再拉长",deliverables:"in.classical_md、提交脚本、dump 轨迹和 thermo 日志"},expectedOutputs:"in.classical_md、run_lammps.slurm、dump 轨迹、log 文件和后处理入口",promptTemplate:["请为 {{system}} 体系配置 LAMMPS 经典力场分子动力学。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请按 EAM / ReaxFF / LJ 等力场生成输入脚本，检查单位与边界，并给出提交与后处理计划。"].join(`
-`)},{id:"lammps-postprocess",title:"轨迹后处理",summary:"分析 LAMMPS 轨迹、能量、RDF、MSD 等结果",fields:i,capabilitySummary:{current:"LAMMPS 输出后处理",keyConfig:"RDF、MSD、能量/温度稳定性",resources:"本地或登录节点即可处理中等轨迹",deliverables:"RDF/MSD 图、能量曲线、统计表和结果说明"},expectedOutputs:"RDF/MSD 图、能量曲线、统计表和结果说明",promptTemplate:["请分析 {{system}} 体系的 LAMMPS 轨迹与热力学输出。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请生成后处理图表，并给出模拟是否稳定、是否需要继续采样的结论。"].join(`
-`)}]},_e={id:"vasp",name:"VASP",category:"hpc",domain:"materials",tags:["HPC","DFT","AIMD"],blurb:"第一性原理材料模拟软件，用于结构优化、静态能量、能带/DOS 和 AIMD 数据生成。",docsHref:"https://gitee.com/onescience-ai/onescience/tree/main/examples/matchem/tools/vasp",intro:{title:"30 秒看懂 HPC 软件",subtitle:"用于 DFT 结构优化、静态能量、能带、DOS 和 AIMD 计算，是材料模拟流程中的高精度计算底座",suited:["结构优化","DFT 静态计算","AIMD 数据生成","能带 / DOS"],notFor:["没有 VASP 许可证","POTCAR 不完整","体系太大且只需快速筛选"],prepare:["POSCAR","INCAR","KPOINTS","POTCAR","Slurm 脚本"],obtain:["OUTCAR","CONTCAR","XDATCAR","vasprun.xml","能量和力数据"],steps:["检查许可证和模块","检查四件套输入","生成提交脚本","提交 VASP 作业","解析结果并生成后处理任务"]},install:{banner:"已为你创建 VASP 软件环境配置任务",prompt:["请检查当前项目和远程 HPC 环境是否可以运行 VASP，并完成软件环境配置方案。","软件定位：第一性原理材料模拟软件。","VASP 版本：intel2021+CPVASP+VASPSol+vtst+固定基矢优化","下载链接：curl -f -C - -o dcu-vasp642-25Aug2026-dtk2604-all.tgz https://zzefile.scnet.cn:65011/efile/s/d/bGl1amllMjg2MA==/ad422d07bc4c758d","典型输入：POSCAR / INCAR / KPOINTS / POTCAR / Slurm 脚本。","典型输出：OUTCAR / CONTCAR / XDATCAR / vasprun.xml / 能量和力数据。","要求：","1. 先检查 module、conda/container、MPI/GPU、许可证/访问权限和项目目录权限。","2. 说明 VASP 适用的运行方式、推荐版本、依赖环境和验证命令。","3. 如果需要安装依赖、加载模块或提交远程命令，请先向我请求批准。","4. 安装/配置完成后，请给出最小验证用例，并说明后续可以做：结构优化计算、AIMD 数据生成、结果检查与后处理。"].join(`
-`)},scenarios:[{id:"vasp-relax",title:"结构优化计算",summary:"检查 POSCAR/INCAR/KPOINTS/POTCAR 并提交 relax 任务",fields:i,capabilitySummary:{current:"VASP DFT 结构优化",keyConfig:"ENCUT、EDIFF、ISIF、KPOINTS、力收敛",resources:"HPC MPI 队列；需有效许可证和 POTCAR",deliverables:"可运行目录、任务 ID、CONTCAR 和收敛报告"},expectedOutputs:"可运行 VASP 目录、run_vasp.slurm、任务 ID、CONTCAR 和收敛报告",promptTemplate:["请为 {{system}} 体系准备并提交 VASP 结构优化计算。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请先检查四件套输入和许可证，再生成提交脚本并说明收敛判据。"].join(`
-`)},{id:"vasp-aimd",title:"AIMD 数据生成",summary:"生成 AIMD 输入文件，用于后续 MACE/DeePMD 训练数据",fields:i,capabilitySummary:{current:"VASP AIMD 数据生成",keyConfig:"IBRION=0、TEBEG/TEEND、POTIM、NSW",resources:"HPC 队列；轨迹用于 MACE/DeePMD 训练",deliverables:"AIMD INCAR、提交脚本、XDATCAR/OUTCAR 路径和数据整理建议"},expectedOutputs:"AIMD INCAR、提交脚本、XDATCAR/OUTCAR 路径和数据整理建议",promptTemplate:["请为 {{system}} 体系准备 VASP AIMD 计算，用于机器学习势训练数据。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请给出温度、步长、总步数建议，并说明如何把轨迹转为 MACE/DeePMD 数据集。"].join(`
-`)},{id:"vasp-postprocess",title:"结果检查与后处理",summary:"检查收敛、能量、力和轨迹，并生成后处理摘要",fields:i,capabilitySummary:{current:"VASP 结果检查与后处理",keyConfig:"收敛判据、能量/力趋势、轨迹完整性",resources:"本地或登录节点即可处理中等输出",deliverables:"收敛报告、能量/力曲线、轨迹摘要和后续建议"},expectedOutputs:"收敛报告、能量/力曲线、轨迹摘要和后续建议",promptTemplate:["请检查 {{system}} 体系的 VASP 计算结果并做后处理摘要。","输入数据：{{inputData}}","任务目标：{{goal}}","运行环境：{{environment}}","请确认收敛、能量、力和轨迹是否可用，并给出后处理图表与下一步建议。"].join(`
-`)}]},s=[me,ce,...de,ue,_e],he=[{id:"onemodel",label:"onemodel",title:"模型库",copy:"选择科研基座模型，安装后可直接进入对话工作台使用"},{id:"hpc",label:"HPC 软件",title:"软件库",copy:"按领域管理 LAMMPS、VASP 等科研计算软件，连接远程 HPC 运行环境"},{id:"onedata",label:"onedata",title:"数据资产",copy:"管理公开数据集、数据契约与科研数据接入能力，后续将在此展示数据卡片"},{id:"oneskills",label:"oneskills",title:"技能编排",copy:"按编排层、资源层、专家层与执行层浏览 OneSkills 官方技能，并支持一键更新"}];function fe(n){return he.map(r=>({...r,items:s.filter(o=>o.category===r.id&&(!n||o.domain===n))}))}function ye(n){return s.find(r=>r.id===n)}export{fe as c,ye as g};