@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.040_graniteguardian.zh.yaml

@@ -0,0 +1,174 @@
+id: aisb.t3.040_graniteguardian
+name: '花岗岩卫士：全面的LLM安全防护'
+version: 0.1.0
+one_line: 多头LLM安全防护，可在配置的标准下检测越狱、幻觉及其他跨提示和响应的风险。
+task_description: '本封装基准测试涵盖Granite Guardian风格的LLM安全防护，涉及风险维度包括越狱、社交偏见、粗口、暴力、色情内容、不道德行为以及RAG特定的幻觉风险（上下文相关性、扎根性、答案相关性）。任务涉及评估或校准风险检测头，以提升安全检测指标。当前快照捆绑了cookbook风格的使用路径和Jupyter笔记本，演示使用vLLM进行模型推理，但用于忠实复现指标的XSTEST风格评估器尚未包含在归档中。在将此作为完整基准测试路径之前，需要针对完整指标契约进行外部评估。
+
+  '
+capability_tags:
+- llm_safety
+- guardrails
+- risk_detection
+- classification
+- evaluation
+- rag_quality
+- content_moderation
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- xstest_rh_auc
+- xstest_rh_f1
+- xstest_rr_auc
+- xstest_rr_f1
+- risk_scoring_examples
+- guardrail_auc
+- rag_hallucination_auc
+launch_profiles:
+- id: cookbook_demo
+  label: Cookbook演示
+  description: '运行封装的Granite Guardian cookbook（基于vLLM的笔记本），以验证配置标准下的模型使用和风险评分行为。涵盖越狱检测、RAG扎根性检查和自定义标准工作流。
+
+    '
+- id: restore_eval
+  label: 恢复评估路径
+  description: '在将此基准测试作为忠实指标复现路径之前，请附加或恢复缺失的XSTEST风格评估器。当前快照缺少XSTEST指标的可执行锚点。
+
+    '
+- id: external_benchmark_eval
+  label: 外部基准测试评估
+  description: '在外部基准测试上评估Granite Guardian模型，包括XSTEST、LLM-AggreFact和REVEAL，以获取可复现的AUC和F1分数。
+
+    '
+dataset_download:
+  primary_method: huggingface
+  sources:
+  - name: Granite Guardian模型
+    url: https://huggingface.co/ibm-granite
+    notes:
+    - granite-guardian-3.3-8b
+    - granite-guardian-3.2-5b
+    - granite-guardian-3.2-3b-a800m
+    - granite-guardian-3.1-8b
+    - granite-guardian-3.1-2b
+  - name: 评估数据集
+    url: https://huggingface.co/ibm
+    notes:
+    - SocialStigmaQA
+    - AttaQ
+    - ProvoQ
+    - WikiContradict
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 所有模型权重和数据集均可公开访问
+  - HF Spaces演示地址：https://huggingface.co/spaces/ibm-granite/granite-guardian
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 48
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torch
+  - transformers
+  - vllm
+  notes:
+  - 参见捆绑的README/requirements获取完整依赖集
+  - 推荐使用vLLM进行高效推理
+  - 内容警告：示例可能包含冒犯性语言、刻板印象或歧视性内容
+risk_flags:
+- blocked_metrics
+- partial_snapshot
+risk_notes:
+- xstest_rh_auc指标在当前快照中没有可执行代码锚点
+- xstest_rh_f1指标在当前快照中没有可执行代码锚点
+- xstest_rr_auc指标在当前快照中没有可执行代码锚点
+- xstest_rr_f1指标在当前快照中没有可执行代码锚点
+- 本次打包未执行基准测试
+- 在恢复源文件或获取更完整的本地快照之前，请勿使用
+- 恢复规则：选择origin_path、source_ref和evaluation_protocol.code_paths时，优先使用现有代码文件而非仅README声明
+recommended_when: '当您需要带有可配置标准的实用LLM安全防护评估任务、生产级监控用例或RAG管道安全验证时使用此基准测试。特别适合评估检索增强生成场景中的风险检测（扎根性、上下文相关性、答案相关性）以及越狱抵抗测试。
+
+  '
+not_recommended_when: '如果您需要当前快照中具有XSTEST指标锚点的完全可执行基准测试，或者您的重点是生成质量而非安全和验证，请勿使用。当需要可复现指标而无需外部评估设置时也不适用。
+
+  '
+paper:
+  title: 'Granite Guardian: Comprehensive LLM Safeguarding'
+  authors: 'Inkit Padhi, Manish Nagireddy, Giandomenico Cornacchia, Subhajit Chaudhury,
+    Tejaswini Pedapati, Pierre Dognin, Keerthiram Murugesan, Erik Miehling, Martin
+    Santillan, Cooper Kieran Fraser, Giulio Zizzo, Muhammad Zaid Hameed, Mark Purcell,
+    Michael Desmond, Qian Pan, Zahra Ashktorab, Inge Vejsbjerg, Elizabeth Daly, Michael
+    Hind, Werner Geyer, Ambrish Rawat, Kush R. Varshney, Prasanna Sattigeri
+
+    '
+  venue: NAACL 2025 Industry Track
+  year: 2025
+  url: https://arxiv.org/abs/2412.07724
+  github: https://github.com/ibm-granite/granite-guardian
+  license: Apache 2.0
+  key_findings:
+  - 在有害内容基准测试上达到0.871的AUC分数
+  - 在RAG幻觉相关基准测试上达到0.854的AUC分数
+  - 在部署导向指标上优于其他开源和闭源模型
+  - 在结合人工标注和合成数据的独特数据集上训练
+  - 支持自定义标准用于自定义风险检测
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.040_graniteguardian.zip
+  archive_type: zip
+  local_dir_name: paper-40-GraniteGuardian
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.040_graniteguardian.zip
+  sha256: c98d57a175b4dc64b7f4bb81e3c4c16adb38a98775335f077f780d491e92b4e2
+  size_bytes: 2545921
+display:
+  palette_seed: indigo-sand-shield
+  art_style: safety-console
+  accent_priority: high
+image_path: ../image/040_aisb.t3.040_graniteguardian.jpg
+code_anchors:
+  executable:
+  - cookbooks/granite-guardian-3.3/quickstart.ipynb
+  - cookbooks/granite-guardian-3.2/detailed_guide_vllm.ipynb
+  - cookbooks/granite-guardian-3.1/quick_start_vllm.ipynb
+  code_roots:
+  - cookbooks/
+  blocked_metrics:
+  - xstest_rh_auc
+  - xstest_rh_f1
+  - xstest_rr_auc
+  - xstest_rr_f1
+recovery_rule: '选择origin_path、source_ref和evaluation_protocol.code_paths时，优先使用现有代码文件而非仅README声明。在获得可信执行输出之前，保持metrics_summary为空。如果当前快照中不存在可执行锚点，则将指标标记为临时/阻止状态。
+
+  '
+commercial:
+  annual_fee: null
+  notes: 模型以Apache 2.0许可证开源发布

package/AISB/catalog/aisb.t3.041_amdm.yaml

@@ -0,0 +1,149 @@
+id: aisb.t3.041_amdm
+name: Auto-Regressive Moving Diffusion Models for Time Series Forecasting
+version: 0.1.0
+one_line: Diffusion-based time-series forecasting with auto-regressive moving structure
+  and continuous sequential diffusion for improved probabilistic forecasting.
+task_description: 'This benchmark implements the Auto-Regressive Moving Diffusion
+  (ARMD) model for time series forecasting. Unlike traditional diffusion models that
+  start from white Gaussian noise, ARMD employs chain-based diffusion with priors,
+  modeling the evolution of time series as a continuous sequential diffusion process.
+  The forward diffusion diffuses the future series into the historical series, while
+  the reverse denoising/devolution process transforms historical series into future
+  forecasts. This aligns the diffusion sampling procedure with the forecasting objective,
+  removing the need for conditional generation. The task is to tune diffusion schedules
+  and moving dynamics to improve probabilistic forecasting metrics on standard public
+  benchmark suites.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- forecast_samples
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Inspect the preserved diffusion forecasting code and configs before
+    a full run.
+- id: restore_first
+  label: Restore First
+  description: Restore the missing main entry files referenced by the README before
+    treating this package as runnable.
+- id: full_train_eval
+  label: Full Train and Evaluate
+  description: Run training with scheduler-aware training and evaluate on all metrics.
+    Use config files in Config/ directory (etth1.yaml, ettm1.yaml, exchange.yaml,
+    stock.yaml).
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: ETT dataset
+    url: https://github.com/thuml/iTransformer
+    license: MIT
+  - name: Solar Energy and Exchange datasets
+    url: https://github.com/thuml/iTransformer
+    license: MIT
+  - name: Stock dataset
+    url: https://github.com/Y-debug-sys/Diffusion-TS
+    license: MIT
+  notes:
+  - Download datasets and place in ./Data/datasets directory
+  - Datasets must be obtained from external repositories as they are not bundled
+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
+  flash_attn: null
+  key_packages:
+  - ema_pytorch
+  - tqdm
+  - torch
+  - numpy
+  - scipy
+  - scikit-learn
+  - matplotlib
+  - seaborn
+  notes:
+  - Install dependencies via: pip install -r requirements.txt
+  - Full dependency set documented in bundled README
+risk_flags:
+- partial_snapshot
+- no_runtime_verification
+risk_notes:
+- Static code audit found executable anchors for all staged metrics
+- No benchmark execution was performed in this packaging pass
+- Runtime execution required before trusting metric values
+- Recovery mode recommended to restore any missing main entry files
+recommended_when: 'Use this benchmark when you need a diffusion-style probabilistic
+  forecasting task with standard public time-series datasets, require state-of-the-art
+  performance on multivariate forecasting, or want to explore continuous sequential
+  diffusion methods that align diffusion sampling with forecasting objectives.
+
+  '
+not_recommended_when: 'Do not use this if you need a tiny CPU-only benchmark, require
+  a non-generative point-estimate forecasting setup, or need pre-executed and verified
+  metric values without runtime overhead.
+
+  '
+paper:
+  title: Auto-Regressive Moving Diffusion Models for Time Series Forecasting
+  authors:
+  - Jiaxin Gao
+  - Qinglong Cao
+  - Yuntian Chen
+  venue: arXiv preprint
+  year: 2024
+  url: https://arxiv.org/abs/2412.09328
+  code_url: https://github.com/daxin007/ARMD
+display:
+  palette_seed: midnight-amber-diffusion
+  art_style: sequential-lab
+  accent_priority: high
+image_path: ../image/041_aisb.t3.041_amdm.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- diffusion_models
+- forecasting
+- scientific_ml
+- probabilistic_forecasting
+- autoregressive_moving
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.041_amdm.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.041_amdm.zip
+  archive_type: zip
+  sha256: 9fc41b0433a71fc79ca141b9d6451c7323ab4286cead7d38e77d5ed793283244
+  size_bytes: 2172119

package/AISB/catalog/aisb.t3.041_amdm.zh.yaml

@@ -0,0 +1,137 @@
+id: aisb.t3.041_amdm
+name: 时间序列预测的自回归移动扩散模型
+version: 0.1.0
+one_line: 基于扩散的时间序列预测，采用自回归移动结构和连续顺序扩散，提升概率预测能力。
+task_description: >
+  本基准实现了用于时间序列预测的自回归移动扩散（ARMD）模型。与从白高斯噪声开始
+  的传统扩散模型不同，ARMD采用基于链的扩散与先验，将时间序列的演化建模为连续顺序扩
+  散过程。前向扩散将未来序列扩散到历史序列，而反向去噪/演化过程则将历史序列转换为
+  未来预测。这使扩散采样过程与预测目标对齐，无需条件生成。任务是通过调整扩散调度
+  和移动动态来提升标准公开基准套件上的概率预测指标。
+
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- forecast_samples
+launch_profiles:
+- id: analysis_only
+  label: 仅分析
+  description: 在完整运行前，检查保存的扩散预测代码和配置。
+- id: restore_first
+  label: 优先恢复
+  description: 在将此包视为可运行之前，恢复README中引用的缺失主入口文件。
+- id: full_train_eval
+  label: 完整训练和评估
+  description: 使用调度器感知训练进行训练，并在所有指标上评估。使用Config/目录下的
+    配置文件（etth1.yaml、ettm1.yaml、exchange.yaml、stock.yaml）。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: ETT数据集
+    url: https://github.com/thuml/iTransformer
+    license: MIT
+  - name: 太阳能和汇率数据集
+    url: https://github.com/thuml/iTransformer
+    license: MIT
+  - name: 股票数据集
+    url: https://github.com/Y-debug-sys/Diffusion-TS
+    license: MIT
+  notes:
+  - 下载数据集并放置在 ./Data/datasets 目录
+  - 数据集必须从外部仓库获取，因未捆绑
+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
+  flash_attn: null
+  key_packages:
+  - ema_pytorch
+  - tqdm
+  - torch
+  - numpy
+  - scipy
+  - scikit-learn
+  - matplotlib
+  - seaborn
+  notes:
+  - 通过以下命令安装依赖：pip install -r requirements.txt
+  - 完整依赖集记录在捆绑的README中
+risk_flags:
+- partial_snapshot
+- no_runtime_verification
+risk_notes:
+- 静态代码审计发现所有分阶段指标都有可执行锚点
+- 本次打包未执行基准测试
+- 运行时执行是验证指标值之前必需的
+- 建议使用恢复模式以恢复任何缺失的主入口文件
+recommended_when: >
+  当您需要使用标准公开时间序列数据集进行扩散式概率预测任务、需要在多元预测上达
+  到最优性能，或想要探索将扩散采样与预测目标对齐的连续顺序扩散方法时，请使用此基准。
+
+not_recommended_when: >
+  如果您需要仅CPU运行的小型基准测试、需要非生成式的点估计预测设置，或需要在无运行
+  时开销的情况下获得预先执行并验证的指标值，请勿使用此基准。
+
+paper:
+  title: Auto-Regressive Moving Diffusion Models for Time Series Forecasting
+  authors:
+  - Jiaxin Gao
+  - Qinglong Cao
+  - Yuntian Chen
+  venue: arXiv preprint
+  year: 2024
+  url: https://arxiv.org/abs/2412.09328
+  code_url: https://github.com/daxin007/ARMD
+display:
+  palette_seed: midnight-amber-diffusion
+  art_style: sequential-lab
+  accent_priority: high
+image_path: ../image/041_aisb.t3.041_amdm.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- diffusion_models
+- forecasting
+- scientific_ml
+- probabilistic_forecasting
+- autoregressive_moving
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.041_amdm.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.041_amdm.zip
+  archive_type: zip
+  sha256: 9fc41b0433a71fc79ca141b9d6451c7323ab4286cead7d38e77d5ed793283244
+  size_bytes: 2172119

package/AISB/catalog/aisb.t3.042_xpatch.yaml

@@ -0,0 +1,216 @@
+id: aisb.t3.042_xpatch
+name: 'xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend
+  Decomposition'
+version: 0.1.0
+one_line: CNN-MLP dual-stream long-term time-series forecasting with EMA-based seasonal-trend
+  decomposition and long-context patching.
+task_description: 'This packaged benchmark covers xPatch, a dual-stream long-term
+  time series forecasting (LTSF) architecture that decomposes input series into seasonal
+  and trend components using Exponential Moving Average (EMA) rather than simple moving
+  average. The linear stream employs MLP layers for trend prediction while the nonlinear
+  stream uses depthwise separable CNN for seasonal patterns. The model investigates
+  patching and channel-independence techniques within a non-transformer backbone,
+  evaluating against transformer baselines (PatchTST, Autoformer, FEDformer, CARD)
+  and linear baselines (DLinear, RLinear) on standard LTSF benchmark datasets.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- forecast_checkpoints
+- ablation_outputs
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run xPatch on a single prepared forecasting dataset using unified
+    experimental settings. Executes xPatch_unified script with default hyperparameters.
+    Suitable for rapid validation of model behavior on one benchmark dataset.
+
+    '
+- id: hyper_search
+  label: Hyperparameter Search
+  description: 'Run comprehensive hyperparameter search across patch length, lookback
+    windows, prediction lengths, and EMA alpha parameter. Uses xPatch_search script.
+    Produces results for fair comparison against baseline papers that report with
+    drop-last trick.
+
+    '
+- id: fair_benchmark
+  label: Fair Benchmark (TFB/OpenTS)
+  description: 'Execute fair experiments without the drop-last trick following the
+    TFB benchmark protocol. Uses xPatch_fair script. Results are comparable to OpenTS
+    leaderboard entries.
+
+    '
+- id: ablation_suite
+  label: Ablation Suite
+  description: 'Run ablation studies from the paper appendix including EMA decomposition
+    impact, arctangent loss effect, sigmoid learning rate adjustment scheme, and inference
+    time comparisons against baseline models in ./ablation directory.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: Google Drive
+    url: https://drive.google.com/u/0/uc?id=1NF7VEefXCmXuWNbnNe858WvQAkJ_7wuP&export=download
+    format: csv
+  - name: Baidu Pan
+    url: https://pan.baidu.com/share/init?surl=r3KhGd0Q9PJIUZdfEYoymg&pwd=i9iy
+    format: csv
+  - name: Kaggle Datasets
+    url: https://www.kaggle.com/datasets/wentixxiaogege/time-series-dataset
+    format: csv
+  notes:
+  - Download all preprocessed datasets into ./dataset directory
+  - Benchmark datasets include: ETTh1, ETTh2, ETTm1, ETTm2, Weather, Traffic, Electricity
+  - Datasets are pre-processed and ready for direct use without additional normalization
+    steps
+credential_requirements:
+  mode: optional
+  items:
+  - kaggle_api_token
+  notes:
+  - Kaggle download requires kaggle credentials or API token
+  - Google Drive and Baidu links available without authentication
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: Single GPU sufficient; RTX 6000-class or equivalent recommended
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 'Original experiments used single Quadro RTX 6000 (24GB). Hyperparameter
+      search and ablation studies benefit from additional RAM for larger batch sizes.
+
+      '
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - pytorch>=2.0.0
+  - numpy
+  - pandas
+  - scikit-learn
+  - matplotlib
+  notes:
+  - Environment file (environment.yml) specifies conda CUDA 11.8 toolkit with pip-installed
+    PyTorch 2.5.1
+  - Mixed conda/pip stack requires careful environment management
+  - Create conda environment with: conda env create -f environment.yml
+  - See bundled README for full dependency set
+  - Experiment scripts in ./scripts directory handle runtime configuration
+risk_flags:
+- mixed_environment_stack
+- data_download_external
+- hyperparameter_sensitivity
+risk_notes:
+- Mixed conda CUDA 11.8 with pip PyTorch 2.5.1 may cause version conflicts; verify
+  CUDA runtime matches PyTorch built-in CUDA
+- External dataset downloads from Google Drive/Baidu may require VPN or regional access
+- Performance is sensitive to patch length, lookback window, prediction length, and
+  EMA alpha parameter
+- Fair benchmark results exclude drop-last trick per TFB protocol; verify which experimental
+  setting matches your evaluation goals
+recommended_when: 'Use this benchmark when evaluating dual-stream non-transformer
+  forecasting architectures, studying EMA-based seasonal-trend decomposition effects,
+  investigating patching and channel-independence outside transformer backbones, or
+  comparing CNN/MLP approaches against established transformer and linear baselines
+  on long-term forecasting tasks.
+
+  '
+not_recommended_when: 'Do not use this benchmark for non-time-series tasks, real-time/online
+  forecasting scenarios, scenarios without access to workstation-class GPU with 24GB
+  VRAM, or when requiring models with built-in uncertainty quantification.
+
+  '
+paper:
+  title: 'xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend
+    Decomposition'
+  authors:
+  - Artyom Stitsyuk
+  - Jaesik Choi
+  venue: AAAI 2025
+  year: 2025
+  volume: '39'
+  number: '19'
+  pages: 20601-20609
+  url: https://ojs.aaai.org/index.php/AAAI/article/view/34270
+  arxiv_url: https://arxiv.org/pdf/2412.17323
+  bibtex: "@inproceedings{stitsyuk2025xpatch,\n  title={xPatch: Dual-Stream Time Series\
+    \ Forecasting with Exponential Seasonal-Trend Decomposition},\n  author={Stitsyuk,\
+    \ Artyom and Choi, Jaesik},\n  booktitle={Proceedings of the AAAI Conference on\
+    \ Artificial Intelligence},\n  volume={39}, number={19}, pages={20601--20609},\
+    \ year={2025}\n}\n"
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.042_xpatch.zip
+  archive_type: zip
+  local_dir_name: paper-42-xPatch
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.042_xpatch.zip
+  sha256: 72c8e6fd7e1a8b2eb184077ce9a37fb36514ef7bb148d8d0ff4ed79d4aa9de25
+  size_bytes: 3621184
+display:
+  palette_seed: teal-sunrise-trend
+  art_style: signal-geometry
+  accent_priority: high
+image_path: ../image/042_aisb.t3.042_xpatch.jpg
+metric_contract:
+  mse:
+    origin_path: utils/metrics.py
+    source_ref: MSE function
+    evaluation_protocol:
+      code_paths:
+      - run.py
+      - utils/metrics.py
+    status: code_backed
+  mae:
+    origin_path: utils/metrics.py
+    source_ref: MAE function
+    evaluation_protocol:
+      code_paths:
+      - run.py
+      - utils/metrics.py
+    status: code_backed
+  forecast_checkpoints:
+    origin_path: exp/exp_main.py
+    source_ref: model state saving
+    status: provisional
+code_anchors:
+- environment.yml
+- run.py
+- utils/metrics.py
+- exp/exp_main.py
+- layers/decomp.py
+- layers/network.py
+- layers/network_cnn.py
+- layers/network_mlp.py
+- models/xPatch.py
+caveat_snapshot: 'Static code audit found executable anchors for MSE and MAE metrics
+  in utils/metrics.py and run.py. No benchmark execution was performed in this packaging
+  pass. Runtime execution is required before trusting metric values. Fair benchmark
+  requires TFB/OpenTS protocol compliance; confirm drop-last trick handling matches
+  evaluation goals.
+
+  '