@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.005_tsae.yaml

@@ -0,0 +1,193 @@
+schema_version: 1
+id: aisb.t3.005_tsae
+name: 'Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language
+  for Interpretability'
+version: 0.1.0
+one_line: Train temporal sparse autoencoders (T-SAEs) with a contrastive temporal-consistency
+  loss on LLM activations, then evaluate feature quality via SAEBench-style automated
+  interpretability scoring.
+task_description: 'This benchmark requires training Temporal Sparse Autoencoders (T-SAEs)
+  that partition latent features into high-level (semantic, temporally consistent)
+  and low-level (syntactic, token-local) components using a contrastive loss over
+  adjacent-token activations from language models (Pythia-160m or Gemma-2-2b). The
+  T-SAE architecture extends Matryoshka BatchTopK SAEs with a temporal contrastive
+  term encouraging high-level features to remain stable across adjacent tokens in
+  a sequence. The primary execution route is: (1) train a T-SAE on model activations
+  extracted from the Pile dataset using dictionary_learning/dictionary_learning/train_temporal.py,
+  (2) evaluate reconstruction quality, sparsity, and temporal smoothness using dictionary_learning/dictionary_learning/eval_temporal.py,
+  and (3) compute automated interpretability scores by processing SAEBench-generated
+  explanation JSON files via src/process_sae_bench_autointerp.py. Downstream experiments
+  include probing for semantic/ contextual/syntactic content, t-SNE visualization,
+  and an HH-RLHF alignment case study. Pretrained Gemma-2-2b T-SAE checkpoints with
+  16384 features and pre-labeled explanations are available on HuggingFace, enabling
+  evaluation without retraining. The autointerp evaluation requires SAEBench (external)
+  to generate feature explanations using a judge LLM (Llama-3.3-70b-Instruct in the
+  paper). Training data streams from the Pile via HuggingFace datasets; no manual
+  dataset download is needed for the core route.
+
+  '
+capability_tags:
+- research_code_optimization
+- mechanistic_interpretability
+- sparse_autoencoders
+- large_language_models
+- evaluation
+- contrastive_learning
+- dictionary_learning
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 1d+
+difficulty: hard
+data_access: public
+primary_outputs:
+- autointerp_score
+- sae_checkpoint
+- interpretability_report
+- probing_accuracy
+- temporal_smoothness_metrics
+launch_profiles:
+- id: quick_check
+  label: Quick Check
+  description: 'Run a small T-SAE training pass on Pythia-160m with reduced steps
+    and a short eval to verify the pipeline end-to-end. Uses NeelNanda/pile-10k for
+    fast iteration.
+
+    '
+- id: autointerp_eval
+  label: Autointerp Eval
+  description: 'Train temporal sparse autoencoders on full Pile data, generate feature
+    explanations via SAEBench with a judge LLM, and compute autointerp scores. This
+    is the paper-faithful route requiring external SAEBench and a 70B judge model.
+
+    '
+- id: pretrained_eval
+  label: Pretrained Eval
+  description: 'Download the pretrained Gemma-2-2b T-SAE and bundled explanations
+    from HuggingFace, then run eval_temporal.py and process_sae_bench_autointerp.py
+    without retraining.
+
+    '
+dataset_download:
+  primary_method: streaming
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/datasets/monology/pile-uncopyrighted
+    access: public
+    note: Primary training/eval data; streamed via HuggingFace datasets. No manual
+      download required.
+  - kind: huggingface
+    url: https://huggingface.co/datasets/NeelNanda/pile-10k
+    access: public
+    note: Small subset used for quick evaluation passes.
+  - kind: huggingface
+    url: https://huggingface.co/alex-oesterling/temporal-saes
+    access: public
+    note: Pretrained Gemma-2-2b T-SAE checkpoint (16384 features) with labeled explanations.json.
+  notes:
+  - Training data is streamed; disk usage is dominated by model weights and SAE checkpoints.
+  - Gemma-2-2b weights require HuggingFace access agreement.
+credential_requirements:
+  mode: optional
+  items:
+  - HuggingFace token (required for gated models like Gemma-2-2b; optional for Pythia-160m)
+  - SAEBench access (external tool for generating feature explanations; not bundled)
+  notes:
+  - Rename example.env to .env and add HF_TOKEN if using Gemma-2-2b or customizing
+    HuggingFace paths.
+  - Autointerp evaluation requires running SAEBench externally with a judge LLM (paper
+    uses Llama-3.3-70b-Instruct).
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 48
+environment:
+  python: '3.11'
+  cuda: '12.8'
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - poetry
+  - nnsight
+  - datasets
+  - dictionary_learning (bundled fork)
+  notes:
+  - Poetry-managed environment; run 'poetry install' to set up. Torch wheels are pulled
+    from the cu128 index.
+  - The dictionary_learning directory is a bundled fork of github.com/saprmarks/dictionary_learning.
+  - See pyproject.toml and poetry.lock for the full pinned dependency set.
+risk_flags:
+- external_eval_dependency
+- large_model_weights
+- judge_model_required
+- some_experiments_under_construction
+risk_notes:
+- SAEBench is not bundled; autointerp evaluation requires cloning and running it separately
+  with a large judge LLM (70B parameter Llama-3.3 in the paper).
+- Gemma-2-2b weights (~5 GB) must be downloaded and may require HuggingFace access
+  agreement.
+- The full autointerp route requires hosting or API access to a 70B judge model, which
+  significantly increases cost.
+- README marks "Sequence Interpretability (Figures 1 and 4)" and "Steering" experiments
+  as under construction.
+- No benchmark execution was performed during the packaging pass; metrics are code-backed
+  but unverified at runtime.
+recommended_when: 'Use this benchmark when you want a mechanistic-interpretability
+  task that combines SAE training with temporal-contrastive learning and automated
+  interpretability evaluation. Good fit for research on disentangling semantic vs.
+  syntactic features in LLM representations, or for optimizing SAE training pipelines
+  under structured priors.
+
+  '
+not_recommended_when: 'Do not use this if you need a lightweight CPU-only benchmark,
+  cannot host LLM checkpoints (Pythia-160m minimum, Gemma-2-2b recommended) locally,
+  or lack GPU access. Avoid if you cannot run or access a large judge model for the
+  autointerp evaluation route.
+
+  '
+paper:
+  title: 'Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language
+    for Interpretability'
+  authors:
+  - Usha Bhalla
+  - Alex Oesterling
+  - Claudio Mayrink Verdun
+  - Himabindu Lakkaraju
+  - Flavio P. Calmon
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2511.05541
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.005_tsae.zip
+  archive_type: zip
+  local_dir_name: paper-5-TSAE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.005_tsae.zip
+  sha256: f89d8100092417777d135fded098e3fb26a5f38ccdb300658138273924bf070e
+  size_bytes: 322370
+commercial:
+  annual_fee: null
+display:
+  palette_seed: amber-ink-latent
+  art_style: editorial-technical
+  accent_priority: high
+image_path: ../image/005_aisb.t3.005_tsae.jpg

package/AISB/catalog/aisb.t3.005_tsae.zh.yaml

@@ -0,0 +1,139 @@
+schema_version: 1
+id: aisb.t3.005_tsae
+name: 时间稀疏自编码器：利用语言的序列特性实现可解释性
+version: 0.1.0
+one_line: 在LLM激活上训练带有对比时间一致性损失的时间稀疏自编码器(T-SAE)，然后通过SAEBench风格的自动化可解释性评分来评估特征质量。
+task_description: 本基准测试需要训练时间稀疏自编码器(T-SAE)，使用语言模型(Pythia-160m或Gemma-2-2b)相邻token激活上的对比损失，将潜在特征划分为高层级（语义、时间一致）和低层级（句法、token局部）组件。T-SAE架构通过时间对比项扩展了Matryoshka BatchTopK SAE，鼓励高层级特征在序列中的相邻token上保持稳定。主要执行流程为：(1)使用dictionary_learning/dictionary_learning/train_temporal.py从Pile数据集提取的模型激活上训练T-SAE，(2)使用dictionary_learning/dictionary_learning/eval_temporal.py评估重构质量、稀疏性和时间平滑性，(3)通过src/process_sae_bench_autointerp.py处理SAEBench生成的解释JSON文件来计算自动化可解释性评分。下游实验包括语义/上下文/句法内容的探测、t-SNE可视化和HH-RLHF对齐案例研究。预训练的Gemma-2-2b T-SAE检查点（16384特征）和带标签的解释可在HuggingFace上获取，无需重新训练即可评估。自动解释评估需要SAEBench（外部工具）使用评判LLM（论文中使用Llama-3.3-70B-Instruct）生成特征解释。训练数据通过HuggingFace datasets从Pile流式传输；核心流程无需手动下载数据集。
+capability_tags:
+- research_code_optimization
+- mechanistic_interpretability
+- sparse_autoencoders
+- large_language_models
+- evaluation
+- contrastive_learning
+- dictionary_learning
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 1d+
+difficulty: hard
+data_access: public
+primary_outputs:
+- autointerp_score
+- sae_checkpoint
+- interpretability_report
+- probing_accuracy
+- temporal_smoothness_metrics
+launch_profiles:
+- id: quick_check
+  label: 快速检查
+  description: 在Pythia-160m上运行小规模T-SAE训练，使用减少的步数和简短评估来验证端到端流程。使用NeelNanda/pile-10k进行快速迭代。
+- id: autointerp_eval
+  label: 自动解释评估
+  description: 在完整Pile数据上训练时间稀疏自编码器，通过SAEBench和评判LLM生成特征解释，并计算自动解释分数。这是论文原始流程，需要外部SAEBench和70B评判模型。
+- id: pretrained_eval
+  label: 预训练评估
+  description: 从HuggingFace下载预训练的Gemma-2-2b T-SAE和捆绑的解释，然后运行eval_temporal.py和process_sae_bench_autointerp.py，无需重新训练。
+dataset_download:
+  primary_method: streaming
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/datasets/monology/pile-uncopyrighted
+    access: public
+    note: 主要训练/评估数据；通过HuggingFace datasets流式传输。无需手动下载。
+  - kind: huggingface
+    url: https://huggingface.co/datasets/NeelNanda/pile-10k
+    access: public
+    note: 用于快速评估的小规模子集。
+  - kind: huggingface
+    url: https://huggingface.co/alex-oesterling/temporal-saes
+    access: public
+    note: 预训练的Gemma-2-2b T-SAE检查点（16384特征）附带标签化的explanations.json。
+  notes:
+  - 训练数据以流式方式传输；磁盘使用主要由模型权重和SAE检查点占用。
+  - Gemma-2-2b权重需要HuggingFace访问协议。
+credential_requirements:
+  mode: optional
+  items:
+  - HuggingFace令牌（用于Gemma-2-2b等门控模型必需；Pythia-160m可选）
+  - SAEBench访问权限（用于生成特征解释的外部工具；未捆绑）
+  notes:
+  - 如使用Gemma-2-2b或自定义HuggingFace路径，请将example.env重命名为.env并添加HF_TOKEN。
+  - 自动解释评估需要外部运行SAEBench并使用评判LLM（论文使用Llama-3.3-70B-Instruct）。
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 48
+environment:
+  python: '3.11'
+  cuda: '12.8'
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - poetry
+  - nnsight
+  - datasets
+  - dictionary_learning (bundled fork)
+  notes:
+  - Poetry管理的环境；运行'poetry install'进行设置。Torch轮子从cu128索引拉取。
+  - dictionary_learning目录是github.com/saprmarks/dictionary_learning的捆绑fork。
+  - 完整的固定依赖集请参见pyproject.toml和poetry.lock。
+risk_flags:
+- external_eval_dependency
+- large_model_weights
+- judge_model_required
+- some_experiments_under_construction
+risk_notes:
+- SAEBench未捆绑；自动解释评估需要单独克隆和运行它，并使用大型评判LLM（论文中使用70B参数Llama-3.3）。
+- Gemma-2-2b权重（约5 GB）必须下载，可能需要HuggingFace访问协议。
+- 完整的自动解释路线需要托管或API访问70B评判模型，这会显著增加成本。
+- README将"序列可解释性（图1和4）"和"转向"实验标记为建设中。
+- 打包过程中未执行基准测试；指标有代码支持但运行时未验证。
+recommended_when: 当您需要一个结合SAE训练与时间对比学习和自动化可解释性评估的机械可解释性任务时使用此基准测试。非常适合研究LLM表征中语义与句法特征的解纠缠，或在结构化先验下优化SAE训练流程。
+not_recommended_when: 如果您需要轻量级仅CPU基准测试、无法在本地托管LLM检查点（Pythia-160m最小，Gemma-2-2b推荐）或缺乏GPU访问，请勿使用。如果无法运行或访问大型评判模型进行自动解释评估路线，也应避免。
+paper:
+  title: 'Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language
+    for Interpretability'
+  authors:
+  - Usha Bhalla
+  - Alex Oesterling
+  - Claudio Mayrink Verdun
+  - Himabindu Lakkaraju
+  - Flavio P. Calmon
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2511.05541
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.005_tsae.zip
+  archive_type: zip
+  local_dir_name: paper-5-TSAE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.005_tsae.zip
+  sha256: f89d8100092417777d135fded098e3fb26a5f38ccdb300658138273924bf070e
+  size_bytes: 322370
+commercial:
+  annual_fee: null
+display:
+  palette_seed: amber-ink-latent
+  art_style: editorial-technical
+  accent_priority: high
+image_path: ../image/005_aisb.t3.005_tsae.jpg

package/AISB/catalog/aisb.t3.006_physense.yaml

@@ -0,0 +1,194 @@
+schema_version: 1
+id: aisb.t3.006_physense
+name: 'PhySense: Sensor Placement Optimization for Accurate Physics Sensing'
+version: 0.1.0
+one_line: 'Two-stage flow-based physics-field reconstruction and sensor placement
+  optimization across turbulent-flow, sea-temperature, and car-aerodynamics benchmarks.
+
+  '
+task_description: 'This benchmark implements PhySense, a synergistic two-stage framework
+  for physics sensing. Stage 1 trains a flow-matching generative model (DiT for regular
+  grids, Transolver for irregular meshes) with cross-attention to reconstruct dense
+  physical fields from sparse sensor observations at arbitrary placements. Stage 2
+  optimizes sensor positions via projected gradient descent under spatial constraints,
+  using reconstruction feedback to discover informative placements. The agent must
+  run training and inference across three bundled scientific domains: (1) Turbulent-Flow
+  (pipe simulation on regular grid), (2) Sea-Temperature (global SST reanalysis on
+  regular grid with land mask), and (3) Car-Aerodynamics (3D surface pressure on irregular
+  mesh). For each domain the snapshot provides both a base-model inference script
+  (random placement) and a walk inference script (optimized placement). Primary metrics
+  are relative L2 error and MSE loss; secondary output is the optimized sensor layout.
+  Data paths in inference scripts point to external locations that must be populated.
+  Pretrained checkpoints are expected under per-domain ./checkpoints/ directories.
+  No external evaluation service is required; all metrics are computed locally in
+  the inference scripts.
+
+  '
+capability_tags:
+- research_code_optimization
+- physics_ml
+- sensor_placement
+- scientific_ml
+- field_reconstruction
+- flow_matching
+- generative_models
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- relative_l2
+- mse_loss
+- sensor_layout
+launch_profiles:
+- id: quick_check
+  label: Quick Check
+  description: 'Run a single-domain inference route (e.g. pipe_infer.py or sea_infer.py)
+    with a pretrained checkpoint to verify the pipeline produces relative_l2 and mse_loss.
+
+    '
+- id: full_benchmark
+  label: Full Benchmark
+  description: 'Train base reconstruction models and run sensor placement optimization
+    across all three domains (Turbulent-Flow, Sea-Temperature, Car-Aerodynamics),
+    then evaluate both random-placement and optimized-placement reconstruction quality.
+
+    '
+- id: placement_only
+  label: Placement Optimization Only
+  description: 'Starting from pretrained base models, run only the Stage 2 sensor
+    placement optimization (walk scripts) and compare against random-placement baselines.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: github
+    url: https://github.com/thuml/PhySense
+    access: public
+    note: 'Source code repository. Data for each benchmark must be obtained separately;
+      see per-domain README files for download links.
+
+      '
+  notes:
+  - 'Sea-Temperature data is loaded from a hardcoded path (/workspace/mayuezhou/ddim/sea_data_large/)
+    in inference scripts. The agent must download the sea temperature .npy files and
+    update paths accordingly.
+
+    '
+  - 'Turbulent-Flow data is loaded via datasets/senseiver_dataloader.py; the underlying
+    data source must be obtained and placed in the expected location.
+
+    '
+  - 'Car-Aerodynamics uses DDP training scripts; data paths are configured via command-line
+    arguments in shell scripts under Car-Aerodynamics/scripts/.
+
+    '
+  - 'Total dataset size across all three domains is estimated at 20-60 GB depending
+    on resolution choices; exact sizes are not stated in the README.
+
+    '
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  key_packages:
+  - dotwiz
+  - pyyaml
+  - numpy
+  - matplotlib
+  notes:
+  - See requirements.txt for the full dependency set.
+  - 'Inference scripts use torch.cuda.amp.autocast and hardcode GPU device indices
+    (e.g. cuda:5, cuda:0). Adjust device IDs to match your hardware.
+
+    '
+  - 'Car-Aerodynamics uses DDP (DistributedDataParallel); multi-GPU may be needed
+    for training but single-GPU suffices for inference.
+
+    '
+risk_flags:
+- hardcoded_paths
+- hardcoded_device_ids
+- dataset_not_bundled
+- checkpoint_not_verified
+risk_notes:
+- 'Inference scripts contain hardcoded absolute data paths (e.g. /workspace/mayuezhou/ddim/sea_data_large/)
+  and hardcoded CUDA device indices (torch.cuda.set_device(5)). These must be edited
+  before running.
+
+  '
+- 'Pretrained checkpoints (e.g. pipe_best_base_model.pth, sea_best_base_model.pth,
+  pipe_50sensor_opt.pth) are expected under per-domain ./checkpoints/ directories.
+  Availability depends on the snapshot packaging; if missing, training must be run
+  first.
+
+  '
+- 'No benchmark execution was performed during the packaging pass; metric values have
+  not been verified at runtime.
+
+  '
+- 'The walk inference scripts (sensor placement optimization evaluation) require optimized
+  checkpoints that are produced only after Stage 2 training.
+
+  '
+recommended_when: 'Use this benchmark when you want a scientific-ML task that combines
+  physics-field reconstruction with constrained sensor placement optimization, testing
+  generative modeling (flow matching) across heterogeneous physical domains with both
+  regular and irregular geometries.
+
+  '
+not_recommended_when: 'Do not use this if you need a lightweight benchmark with minimal
+  storage, no iterative optimization loop, or no GPU requirement. Also not suitable
+  if you cannot modify hardcoded paths and device indices in the source scripts.
+
+  '
+paper:
+  title: 'PhySense: Sensor Placement Optimization for Accurate Physics Sensing'
+  venue: NeurIPS 2025 Oral
+  year: 2025
+  url: https://arxiv.org/abs/2505.18190
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.006_physense.zip
+  archive_type: zip
+  local_dir_name: paper-6-PhySense
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.006_physense.zip
+  sha256: b35419c897e1c56bad7037fc90249da8589f432ba58831550a257b99d19e8b6c
+  size_bytes: 94438
+commercial:
+  annual_fee: null
+display:
+  palette_seed: slate-cyan-fluid
+  art_style: scientific-visualization
+  accent_priority: high
+image_path: ../image/006_aisb.t3.006_physense.jpg

package/AISB/catalog/aisb.t3.006_physense.zh.yaml

@@ -0,0 +1,118 @@
+schema_version: 1
+id: aisb.t3.006_physense
+name: 'PhySense：用于精确物理感知的传感器布局优化'
+version: 0.1.0
+one_line: '基于双阶段流匹配方法的物理场重建与传感器布局优化，涵盖湍流、海温及汽车空气动力学基准。'
+task_description: '本基准实现了PhySense，一个协同式双阶段物理感知框架。第一阶段训练基于流匹配的条件生成模型（规则网格使用DiT，不规则网格使用Transolver），通过交叉注意力机制从稀疏传感器观测中重建稠密物理场，传感器可放置在任意位置。第二阶段在空间约束下通过投影梯度下降优化传感器位置，利用重建反馈发现信息量最大的放置方案。智能体需在三个捆绑的科学领域运行训练和推理：（1）湍流（规则网格管道模拟），（2）海温（带陆面掩码的规则网格全球海表温度再分析），（3）汽车空气动力学（不规则网格上的3D表面压力）。每个领域提供了基础模型推理脚本（随机放置）和步行推理脚本（优化放置）。主要指标为相对L2误差和MSE损失；次要输出为优化后的传感器布局。推理脚本中的数据路径指向外部位置，需自行填充数据。预训练检查点应存放在各领域的./checkpoints/目录下。无需外部评估服务，所有指标在推理脚本中本地计算。'
+capability_tags:
+- research_code_optimization
+- physics_ml
+- sensor_placement
+- scientific_ml
+- field_reconstruction
+- flow_matching
+- generative_models
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- relative_l2
+- mse_loss
+- sensor_layout
+launch_profiles:
+- id: quick_check
+  label: 快速检查
+  description: '使用预训练检查点运行单领域推理（如pipe_infer.py或sea_infer.py），验证流水线能否输出relative_l2和mse_loss指标。'
+- id: full_benchmark
+  label: 完整基准测试
+  description: '在全部三个领域（湍流、海温、汽车空气动力学）上训练基础重建模型并运行传感器布局优化，然后评估随机放置和优化放置两种方案的重建质量。'
+- id: placement_only
+  label: 仅布局优化
+  description: '基于预训练基础模型，仅运行第二阶段传感器布局优化（步行脚本），并与随机放置基线进行对比。'
+dataset_download:
+  primary_method: mixed
+  sources:
+  - kind: github
+    url: https://github.com/thuml/PhySense
+    access: public
+    note: '源代码仓库。各基准数据需单独获取，详见各领域README文件中的下载链接。'
+  notes:
+  - '海温数据从推理脚本中的硬编码路径（/workspace/mayuezhou/ddim/sea_data_large/）加载。智能体需下载海温.npy文件并相应更新路径。'
+  - '湍流数据通过datasets/senseiver_dataloader.py加载；需获取底层数据源并放置到预期位置。'
+  - '汽车空气动力学使用DDP训练脚本；数据路径通过Car-Aerodynamics/scripts/下shell脚本的命令行参数配置。'
+  - '三个领域的数据集总大小估计为20-60GB，取决于分辨率选择；README中未注明确切大小。'
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  key_packages:
+  - dotwiz
+  - pyyaml
+  - numpy
+  - matplotlib
+  notes:
+  - 完整依赖列表见requirements.txt。
+  - '推理脚本使用torch.cuda.amp.autocast并硬编码GPU设备索引（如cuda:5、cuda:0）。请根据您的硬件调整设备ID。'
+  - '汽车空气动力学使用DDP（DistributedDataParallel）；训练可能需要多GPU，但推理仅需单GPU即可。'
+risk_flags:
+- hardcoded_paths
+- hardcoded_device_ids
+- dataset_not_bundled
+- checkpoint_not_verified
+risk_notes:
+- '推理脚本包含硬编码的绝对数据路径（如/workspace/mayuezhou/ddim/sea_data_large/）和硬编码的CUDA设备索引（torch.cuda.set_device(5)）。运行前必须编辑修改。'
+- '预训练检查点（如pipe_best_base_model.pth、sea_best_base_model.pth、pipe_50sensor_opt.pth）应存放在各领域./checkpoints/目录下。是否可用取决于快照打包情况；如缺失需先运行训练。'
+- '打包过程中未执行基准测试；指标值未经运行时验证。'
+- '步行推理脚本（传感器布局优化评估）需要优化后的检查点，该检查点仅在第二阶段训练后产生。'
+recommended_when: '当您需要一项结合物理场重建与约束传感器布局优化的科学机器学习任务，且该任务需测试跨异构物理领域的生成式建模（流匹配）能力，涉及规则和不规则几何构型时，请使用此基准。'
+not_recommended_when: '如果需要轻量级基准（存储需求小、无迭代优化循环、无GPU要求），请勿使用。此外，如无法修改源代码脚本中的硬编码路径和设备索引，也不适用。'
+paper:
+  title: 'PhySense: Sensor Placement Optimization for Accurate Physics Sensing'
+  venue: NeurIPS 2025 Oral
+  year: 2025
+  url: https://arxiv.org/abs/2505.18190
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.006_physense.zip
+  archive_type: zip
+  local_dir_name: paper-6-PhySense
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.006_physense.zip
+  sha256: b35419c897e1c56bad7037fc90249da8589f432ba58831550a257b99d19e8b6c
+  size_bytes: 94438
+commercial:
+  annual_fee: null
+display:
+  palette_seed: slate-cyan-fluid
+  art_style: scientific-visualization
+  accent_priority: high
+image_path: ../image/006_aisb.t3.006_physense.jpg