@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.068_sempo.yaml

@@ -0,0 +1,187 @@
+id: aisb.t3.068_sempo
+name: 'SEMPO: Lightweight Foundation Models for Time Series Forecasting'
+version: 0.1.0
+one_line: Lightweight 6.5M-parameter time-series foundation model with energy-aware
+  spectral decomposition and mixture-of-prompts experts, pretrainable on a single
+  GPU.
+task_description: 'This packaged benchmark covers SEMPO, a novel lightweight time-series
+  foundation model that achieves strong zero-shot and few-shot forecasting generalization
+  while requiring substantially reduced model size and pretraining scale. SEMPO comprises
+  two key modules: (1) Energy-aware SpEctral Decomposition (EASD), which models both
+  high-energy and low-energy frequency signals—addressing a bias in prior methods
+  that ignore low-energy yet informative spectral components; and (2) Mixture-of-PrOmpts
+  enabled Transformer (MoPFormer), which learns heterogeneous temporal patterns via
+  small dataset-specific prompts and adaptively routes time-series tokens to prompt-based
+  experts for parameter-efficient adaptation. With only 6.5M parameters pretrained
+  on 83M time points, SEMPO surpasses state-of-the-art time-series FMs with hundreds
+  of millions of parameters pretrained on billions of time points, achieving average
+  reductions of 12% and 22% in forecasting errors under zero-shot and few-shot scenarios
+  respectively. The benchmark supports single-GPU pretraining, fine-tuning, and zero-shot
+  evaluation across 16 datasets.
+
+  '
+capability_tags:
+- time_series_forecasting
+- foundation_models
+- lightweight_models
+- spectral_decomposition
+- mixture_of_experts
+- prompt_mixture_models
+- zero_shot_learning
+- few_shot_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
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- rmse
+- mape
+- mspe
+- mse_96
+- mse_192
+- mse_336
+- mse_720
+- forecast_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run zero-shot evaluation on one packaged benchmark route (e.g., ETTh1,
+    weather, electricity). Uses --is_pretraining 0 --is_training 0 --is_zeroshot 1.
+    Downloads and evaluates against pretrained checkpoint from ./checkpoints/.
+
+    '
+  hardware_notes: 'Single GPU with 16GB VRAM minimum; 24GB recommended for comfortable
+    execution.
+
+    '
+- id: pretrain_and_eval
+  label: Pretrain + Eval
+  description: 'Run the full lightweight foundation model workflow: pretrain on UTSD
+    data via sempo_utsd.sh, then evaluate with zero-shot or few-shot scripts.
+
+    '
+  hardware_notes: 'Supports single-GPU and multi-GPU execution on a single node. Single-GPU
+    pretraining is feasible due to compact 6.5M parameter model size. Multi-node distributed
+    training also supported via torch.distributed.
+
+    '
+- id: few_shot_eval
+  label: Few-Shot + Eval
+  description: 'Fine-tune with 5% or 10% dataset fractions using few-shot scripts.
+    Uses --is_pretraining 0 --is_training 1 --is_zeroshot 0.
+
+    '
+  hardware_notes: 'Fine-tuning is lightweight due to MoPFormer''s parameter-efficient
+    prompt-based adaptation.
+
+    '
+dataset_download:
+  primary_method: manual
+  sources:
+  - name: UTSD numpy format
+    url: https://cloud.tsinghua.edu.cn/f/93868e3a9fb144fe9719/
+    note: Primary pretraining dataset format
+  - name: Google Drive
+    url: https://drive.google.com/drive/my-drive?dmr=1&ec=wgc-drive-hero-goto
+    note: Download location for UTSD contents
+  notes:
+  - Place downloaded contents under ./dataset directory
+  - PEMS04 and PEMS07 use a separate format from other UTSD datasets
+  - Pretrained checkpoint is bundled in ./checkpoints/ for evaluation without retraining
+credential_requirements:
+  mode: optional
+  items:
+  - huggingface_api
+  notes:
+  - Required only if evaluating Chronos-Bolt comparison baselines
+  - Chronos pretrained weights from https://huggingface.co/collections/amazon/chronos-models-65f1791d630a8d57cb718444
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+    gpu_arch: any_cuda_11x
+    notes: Single GPU sufficient for pretraining due to 6.5M parameter compact architecture
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_arch: any_cuda_11x
+    notes: 24GB VRAM provides comfortable headroom for larger batch sizes and future
+      experiments
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.2
+  flash_attn: null
+  key_packages:
+  - transformers==4.40.1
+  notes:
+  - Install via pip install -r requirements.txt
+  - PyTorch 2.1.2+cu118 required
+  - See bundled README/requirements.txt for full dependency set
+risk_flags:
+- single_point_of_failure
+- reproducibility_variance
+risk_notes:
+- Seed fixed at 2021 in run.py for reproducibility
+- Evaluation scripts support multiple benchmark datasets; results may vary across
+  domains
+- Chronos-Bolt evaluation requires manual weight download from HuggingFace
+recommended_when: 'Use this benchmark when you want to experiment with a compact foundation-model
+  forecasting approach that supports single-GPU pretraining and demonstrates strong
+  zero-shot/few-shot transfer to diverse time-series domains including energy, transportation,
+  weather, and traffic.
+
+  '
+not_recommended_when: 'Do not use this if your research requires large-cluster pretraining
+  scale as the main difficulty driver, or if you need models with hundreds of millions
+  of parameters.
+
+  '
+paper:
+  title: 'SEMPO: Lightweight Foundation Models for Time Series Forecasting'
+  venue: NeurIPS 2025 Poster
+  year: 2025
+  url: https://arxiv.org/abs/2510.19710
+  code_url: https://github.com/mala-lab/SEMPO
+  authors:
+  - Hui He
+  - Kun Yi
+  - Yuanchi Ma
+  - Qi Zhang
+  - Zhendong Niu
+  - Guansong Pang
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.068_sempo.zip
+  archive_type: zip
+  local_dir_name: paper-68-SEMPO
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.068_sempo.zip
+  sha256: efc7e945bebd2a61b437030d21ca2767951a4620edd5f923cbe73d7f7c457073
+  size_bytes: 485364
+display:
+  palette_seed: ivory-lime-spectrum
+  art_style: lightweight-foundation
+  accent_priority: medium
+commercial:
+  annual_fee: null
+image_path: ../image/068_aisb.t3.068_sempo.jpg

package/AISB/catalog/aisb.t3.068_sempo.zh.yaml

@@ -0,0 +1,148 @@
+id: aisb.t3.068_sempo
+name: 'SEMPO: 轻量级时序预测基础模型'
+version: 0.1.0
+one_line: 6.5M参数轻量级时序基础模型，配备能量感知频谱分解与提示混合专家模块，可在单块GPU上预训练。
+task_description: |
+  本基准测试套件涵盖SEMPO，一种创新的轻量级时序基础模型，在大幅降低模型规模和预训练规模的同时，实现了强大的零样本和少样本预测泛化能力。SEMPO包含两个关键模块：(1) 能量感知频谱分解(EASD)，该模块同时建模高频能量和低频能量信号——解决了先前方法忽略低能量但具有信息价值的频谱成分这一偏差问题；(2) 基于提示混合的Transformer(MoPFormer)，通过小型数据集特定提示学习异构时序模式，并自适应地将时序token路由到基于提示的专家模块，实现参数高效适配。仅用6.5M参数在83M时间点上预训练，SEMPO便超越了具有数亿参数、在数十亿时间点上预训练的最先进时序基础模型，在零样本和少样本场景下分别实现了平均12%和22%的预测误差降低。该基准测试支持单GPU预训练、微调和零样本评估，覆盖16个数据集。
+capability_tags:
+- time_series_forecasting
+- foundation_models
+- lightweight_models
+- spectral_decomposition
+- mixture_of_experts
+- prompt_mixture_models
+- zero_shot_learning
+- few_shot_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
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mse
+- mae
+- rmse
+- mape
+- mspe
+- mse_96
+- mse_192
+- mse_336
+- mse_720
+- forecast_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: |
+    在一个打包的基准测试路线（如ETTh1、weather、electricity）上运行零样本评估。使用参数 --is_pretraining 0 --is_training 0 --is_zeroshot 1。从 ./checkpoints/ 下载并评估预训练checkpoint。
+  hardware_notes: |
+    至少需要16GB显存的单块GPU；建议24GB显存以获得舒适的执行体验。
+- id: pretrain_and_eval
+  label: 预训练 + 评估
+  description: |
+    运行完整的轻量级基础模型工作流：通过 sempo_utsd.sh 在UTSD数据上进行预训练，然后使用零样本或少样本脚本进行评估。
+  hardware_notes: |
+    支持单节点单GPU和多GPU执行。由于紧凑的6.5M参数模型规模，单GPU预训练是可行的。也可通过torch.distributed支持多节点分布式训练。
+- id: few_shot_eval
+  label: 少样本 + 评估
+  description: |
+    使用5%或10%的数据集比例通过少样本脚本进行微调。使用参数 --is_pretraining 0 --is_training 1 --is_zeroshot 0。
+  hardware_notes: |
+    由于MoPFormer的参数高效提示适配机制，微调过程较为轻量。
+dataset_download:
+  primary_method: manual
+  sources:
+  - name: UTSD numpy格式
+    url: https://cloud.tsinghua.edu.cn/f/93868e3a9fb144fe9719/
+    note: 主要预训练数据集格式
+  - name: Google Drive
+    url: https://drive.google.com/drive/my-drive?dmr=1&ec=wgc-drive-hero-goto
+    note: UTSD内容下载位置
+  notes:
+  - 将下载内容放置于 ./dataset 目录下
+  - PEMS04和PEMS07使用与其他UTSD数据集不同的单独格式
+  - 预训练checkpoint已打包在 ./checkpoints/ 中，无需重新训练即可进行评估
+credential_requirements:
+  mode: optional
+  items:
+  - huggingface_api
+  notes:
+  - 仅在评估Chronos-Bolt对比基线时需要
+  - Chronos预训练权重需从 https://huggingface.co/collections/amazon/chronos-models-65f1791d630a8d57cb718444 手动下载
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+    gpu_arch: any_cuda_11x
+    notes: 由于6.5M参数紧凑架构，单GPU足以进行预训练
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+    gpu_arch: any_cuda_11x
+    notes: 24GB显存为更大的批量大小和未来实验提供了充足的空间
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.2
+  flash_attn: null
+  key_packages:
+  - transformers==4.40.1
+  notes:
+  - 通过 pip install -r requirements.txt 安装
+  - 需要PyTorch 2.1.2+cu118
+  - 完整的依赖项列表请参阅附带的README/requirements.txt
+risk_flags:
+- single_point_of_failure
+- reproducibility_variance
+risk_notes:
+- 为保证可重复性，run.py中的随机种子固定为2021
+- 评估脚本支持多个基准数据集；结果可能因领域不同而有所差异
+- Chronos-Bolt评估需要从HuggingFace手动下载权重
+recommended_when: |
+  当您希望实验一种紧凑的基础模型预测方法时使用此基准测试，该方法支持单GPU预训练，并展现出向能源、交通、天气和交通等多样化时序领域的强大零样本/少样本迁移能力。
+not_recommended_when: |
+  如果您的研究需要将大规模集群预训练规模作为主要难度驱动因素，或者您需要具有数亿参数的模型，请勿使用此基准测试。
+paper:
+  title: 'SEMPO: Lightweight Foundation Models for Time Series Forecasting'
+  venue: NeurIPS 2025 Poster
+  year: 2025
+  url: https://arxiv.org/abs/2510.19710
+  code_url: https://github.com/mala-lab/SEMPO
+  authors:
+  - Hui He
+  - Kun Yi
+  - Yuanchi Ma
+  - Qi Zhang
+  - Zhendong Niu
+  - Guansong Pang
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.068_sempo.zip
+  archive_type: zip
+  local_dir_name: paper-68-SEMPO
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.068_sempo.zip
+  sha256: efc7e945bebd2a61b437030d21ca2767951a4620edd5f923cbe73d7f7c457073
+  size_bytes: 485364
+display:
+  palette_seed: ivory-lime-spectrum
+  art_style: lightweight-foundation
+  accent_priority: medium
+commercial:
+  annual_fee: null
+image_path: ../image/068_aisb.t3.068_sempo.jpg

package/AISB/catalog/aisb.t3.069_treehfd.yaml

@@ -0,0 +1,129 @@
+id: aisb.t3.069_treehfd
+name: Tree Ensemble Explainability through the Hoeffding Functional Decomposition
+  and TreeHFD Algorithm
+version: 0.1.0
+one_line: Decompose XGBoost ensemble predictions into interpretable main effects and
+  second-order interactions via the TreeHFD algorithm.
+task_description: 'This packaged benchmark evaluates explainability for XGBoost ensembles
+  using the Hoeffding functional decomposition (HFD). The TreeHFD algorithm decomposes
+  black-box tree-ensemble predictions into a unique sum of lower-dimensional functions—main
+  effects (single variables) and second-order interactions (pairs of variables)—that
+  are intrinsically transparent while preserving model accuracy. Unlike TreeSHAP,
+  the Hoeffding decomposition enforces hierarchical orthogonality, yielding clean
+  main-effect estimates where interactions are properly separated. The benchmark measures
+  decomposition quality via mean squared error against ground-truth analytical targets
+  on a controlled simulation study (p=6 dimensions, n=5000 samples, correlated inputs).
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mse_eta12
+- mse_eta1
+- mse_eta2
+- mse_eta3
+- mse_eta4
+- mse_eta5
+- mse_eta6
+- mse_eta34
+- mse_others
+- functional_decomposition_report
+launch_profiles:
+- id: explainability_eval
+  label: Explainability Eval
+  description: Run the packaged Hoeffding-style decomposition route on the tree-ensemble
+    benchmark. Evaluates decomposition accuracy against analytical ground truth across
+    all main effects and second-order interactions.
+dataset_download:
+  primary_method: pip
+  sources:
+  - url: https://pypi.org/project/treehfd/
+    method: pip
+    package_name: treehfd
+  - url: https://deepscientist.cc/AISB/069_treehfd
+    method: archive
+    archive_type: zip
+    local_dir_name: paper-69-TreeHFD
+  notes:
+  - Install via `pip install treehfd` for latest release
+  - Archive provides full source with examples and tests
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - xgboost>=3.1,<4.0
+  - scipy>=1.16,<2.0
+  - numpy>=1.24
+  - scikit-learn>=1.3
+  notes:
+  - CPU-only execution fully sufficient for all benchmark routes
+  - GPU support optional via xgboost's proprietary NVIDIA modules
+  - Python >=3.11 required due to language features used
+  - Full dependency set in bundled requirements.txt or pyproject.toml
+risk_flags: []
+risk_notes: []
+recommended_when: 'Use this benchmark when you need interpretable explanations for
+  XGBoost models on tabular data with properly separated main effects and interactions.
+  Ideal for CPU-only environments and for comparing against TreeSHAP-based approaches
+  where entanglement of main effects with interactions is a concern.
+
+  '
+not_recommended_when: 'Do not use this if you require GPU acceleration (xgboost GPU
+  mode is unsupported), need explanations for neural networks or non-tree models,
+  or require explanations for high-cardinality categorical variables without one-hot
+  encoding.
+
+  '
+paper:
+  title: Tree Ensemble Explainability through the Hoeffding Functional Decomposition
+    and TreeHFD Algorithm
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=dRLWcpBQxS
+  authors:
+  - Clément Bénard
+  institutions:
+  - Thales cortAIx-Labs
+  - SINCLAIR AI Lab
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.069_treehfd.zip
+  archive_type: zip
+  local_dir_name: paper-69-TreeHFD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.069_treehfd.zip
+  sha256: 7b44fd7e4586498d04dbdb0d6d8059aee6621c4253e2292dda069fab6d4abe37
+  size_bytes: 123618
+display:
+  palette_seed: cedar-ivory-tree
+  art_style: model-decomposition
+  accent_priority: medium
+image_path: ../image/069_aisb.t3.069_treehfd.jpg

package/AISB/catalog/aisb.t3.069_treehfd.zh.yaml

@@ -0,0 +1,111 @@
+id: aisb.t3.069_treehfd
+name: 通过Hoeffding函数分解与TreeHFD算法实现树集成模型可解释性
+version: 0.1.0
+one_line: 通过TreeHFD算法将XGBoost集成预测分解为可解释的主效应和二阶交互作用。
+task_description: |
+  本基准测试包旨在使用Hoeffding函数分解（HFD）评估XGBoost集成的可解释性。TreeHFD算法将黑盒树集成预测分解为唯一的低维函数和——主效应（单变量）和二阶交互作用（变量对）——这些函数本质上是透明的，同时保持了模型精度。与TreeSHAP不同，Hoeffding分解强制执行分层正交性，产生干净的主效应估计，其中交互作用被正确分离。基准测试通过与控制模拟研究（p=6维，n=5000样本，相关输入）中的解析真值对比来测量分解质量。
+
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 1-2h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- mse_eta12
+- mse_eta1
+- mse_eta2
+- mse_eta3
+- mse_eta4
+- mse_eta5
+- mse_eta6
+- mse_eta34
+- mse_others
+- functional_decomposition_report
+launch_profiles:
+- id: explainability_eval
+  label: 可解释性评估
+  description: 在树集成基准上运行打包的Hoeffding风格分解流程。在所有主效应和二阶交互作用上，针对解析真值评估分解精度。
+dataset_download:
+  primary_method: pip
+  sources:
+  - url: https://pypi.org/project/treehfd/
+    method: pip
+    package_name: treehfd
+  - url: https://deepscientist.cc/AISB/069_treehfd
+    method: archive
+    archive_type: zip
+    local_dir_name: paper-69-TreeHFD
+  notes:
+  - 通过 `pip install treehfd` 安装最新版本
+  - 压缩包提供完整源代码及示例和测试
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 4
+    ram_gb: 8
+    disk_gb: 10
+    gpu_count: 0
+    gpu_vram_gb: 0
+  recommended:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+environment:
+  python: '3.11'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - xgboost>=3.1,<4.0
+  - scipy>=1.16,<2.0
+  - numpy>=1.24
+  - scikit-learn>=1.3
+  notes:
+  - CPU仅执行对所有基准测试流程完全足够
+  - GPU支持可选，通过xgboost的专有NVIDIA模块
+  - 由于使用的语言特性，需要Python >=3.11
+  - 完整依赖项在捆绑的requirements.txt或pyproject.toml中
+risk_flags: []
+risk_notes: []
+recommended_when: |
+  当需要对表格数据的XGBoost模型进行可解释性解释、且主效应与交互作用需要正确分离时，使用此基准测试。非常适合CPU仅环境，以及在与TreeSHAP方法进行对比时——当主效应与交互作用的纠缠成为问题时。
+
+not_recommended_when: |
+  如果需要GPU加速（不支持xgboost GPU模式）、需要对神经网络或非树模型进行解释，或者需要对高基数分类变量（未经独热编码）进行解释时，请勿使用此基准测试。
+
+paper:
+  title: Tree Ensemble Explainability through the Hoeffding Functional Decomposition
+    and TreeHFD Algorithm
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=dRLWcpBQxS
+  authors:
+  - Clément Bénard
+  institutions:
+  - Thales cortAIx-Labs
+  - SINCLAIR AI Lab
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.069_treehfd.zip
+  archive_type: zip
+  local_dir_name: paper-69-TreeHFD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.069_treehfd.zip
+  sha256: 7b44fd7e4586498d04dbdb0d6d8059aee6621c4253e2292dda069fab6d4abe37
+  size_bytes: 123618
+display:
+  palette_seed: cedar-ivory-tree
+  art_style: model-decomposition
+  accent_priority: medium
+image_path: ../image/069_aisb.t3.069_treehfd.jpg