@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.052_nsdiff.zh.yaml

@@ -0,0 +1,138 @@
+id: aisb.t3.052_nsdiff
+name: 非平稳扩散概率时间序列预测
+version: 0.1.0
+one_line: 针对非平稳时间序列动力学的基于扩散的概率预测框架，在ETTh基准上通过CRPS进行评估。
+task_description: |
+  NsDiff是一个基于扩散的概率预测理论框架，专为非平稳场景设计。该打包的基准测试包含此方法的训练和评估流程，
+  包括预训练脚本、主实验入口点（./src/experiments/NsDiff.py）以及配置文件。然而，当前的本地快照未保留
+  README中引用的完整可执行源代码树；代码所在目录（configs/、notebooks/）仅提供部分锚点。
+  基准测试需要在执行前进行上游恢复。数据集在执行期间自动下载。评估使用CRPS（连续排序概率得分）作为主要指标。
+
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: 困难
+data_access: public
+primary_outputs:
+- crps
+launch_profiles:
+- id: analysis_only
+  label: 仅分析
+  description: 检查notebooks/uncertainty_variation.ipynb中的保留笔记本、configs/nsdiff.yml中的配置，
+    以及NsDiff路线的捆绑文档。
+- id: restore_first
+  label: 优先恢复
+  description: 在尝试基准测试执行之前，恢复README中引用的缺失源代码树。
+    需要从上游恢复./src/experiments/NsDiff.py和./scripts/目录。
+dataset_download:
+  primary_method: automatic
+  sources:
+  - README指出数据集在执行期间自动下载。
+  - 引用的基准测试包括ETTh1、ETTh2、ETTm1、ETTm2及其他ETT数据集变体。
+  notes:
+  - 将数据集设置视为上游恢复的一部分；当前快照中未捆绑下载脚本。
+  - 执行命令引用dataset_type参数（例如--dataset_type="ETTh1"）。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages:
+  - 完整的依赖项请参见./requirements.txt
+  - 从configs/nsdiff.yml测试配置推断的核心依赖项
+  notes:
+  - 当前的本地包未保留可执行设置文件或除部分configs/和notebooks/根目录外的可运行源代码。
+  - 环境提示待上游恢复后生效。
+  - 安装命令：pip install -r ./requirements.txt
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- crps_metric_blocked
+risk_notes:
+- 当前快照未暴露可执行代码文件，无法锚定CRPS指标。
+- 本次打包过程中未执行任何基准测试；所有指标声明均视为临时性质。
+- configs/nsdiff.yml提供了采样和测试参数（batch_size=256、sampling_size=1000、n_z_samples=200、n_bins=10），
+  但无可执行的评估锚点。
+recommended_when: 当您需要论文支持的规划、上游恢复指导，或对NsDiff扩散预测路线的保留笔记本和配置文件
+  进行探索性分析时使用此条目。
+not_recommended_when: 请勿将本地包视为可直接运行的扩散基准测试。
+  README中引用的源代码树（src/experiments/、scripts/）未在本地保留。
+  在CRPS评估执行之前，需要先上游恢复可执行源代码树。
+paper:
+  title: Non-stationary Diffusion For Probabilistic Time Series Forecasting
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://arxiv.org/abs/2505.08613
+  authors:
+  - Weiwei Ye (equal contribution)
+  - Zhuopeng Xu (equal contribution)
+  - Ning Gui (corresponding author)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.052_nsdiff.zip
+  archive_type: zip
+  local_dir_name: paper-52-NsDiff
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.052_nsdiff.zip
+  sha256: cc513276d021fd69cb749c3d843a64dc99b18428e9d1d66f040d07d62bed139a
+  size_bytes: 435952
+display:
+  palette_seed: navy-mint-diffusion
+  art_style: probabilistic-signal
+  accent_priority: high
+image_path: ../image/052_aisb.t3.052_nsdiff.jpg
+execution_evidence:
+  pretrain_script: bash ./scripts/pretrain_F/ETTh1.sh
+  main_entry: python3 ./src/experiments/NsDiff.py
+  example_args:
+  - --dataset_type=ETTh1
+  - --device=cuda:0
+  - --batch_size=32
+  - --horizon=1
+  - --pred_len=192
+  - --windows=168
+  - --load_pretrain=True
+  notes:
+  - 执行期间必须将PYTHONPATH设置为项目根目录。
+  - 多GPU设置建议使用CUDA_DEVICE_ORDER=PCI_BUS_ID。
+  - seeds参数接受列表格式：--seeds='[1232132, 3]'
+  - --load_pretrain=False选项可用于非预训练运行。
+config_anchors:
+- path: configs/nsdiff.yml
+  role: 主要配置文件
+  contains:
+  - 采样参数：batch_size=256, sampling_size=1000, last_only=True
+  - 测试参数：batch_size=256, n_z_samples=200, n_bins=10, minisample=1
+  - 覆盖率与NLL跟踪参数
+- path: notebooks/uncertainty_variation.ipynb
+  role: 分析笔记本锚点
+- path: configs/csbi.yaml
+  role: 附加配置（非主要）
+- path: configs/tmdm.yml
+  role: 附加配置（非主要）
+code_audit_status: blocked
+metric_blocking_reason: 当前快照中无可用于CRPS评估的可执行代码锚点。
+  configs/nsdiff.yml包含测试参数，但实际评估代码路径未被保留。

package/AISB/catalog/aisb.t3.053_k2vae.yaml

@@ -0,0 +1,158 @@
+schema_version: 1
+id: aisb.t3.053_k2vae
+name: 'K2VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic
+  Time Series Forecasting'
+version: 0.1.0
+one_line: Train and evaluate a Koopman-Kalman VAE for probabilistic time-series forecasting
+  across 17 public datasets, measuring CRPS and NLL.
+task_description: 'This benchmark trains and evaluates K2VAE, a VAE-based generative
+  model that uses a KoopmanNet to linearize nonlinear time series into a latent dynamical
+  system and a KalmanNet to iteratively refine predictions and model process uncertainty
+  via learned Kalman gains. The task covers both short-term forecasting (8 datasets:
+  exchange_rate_nips, electricity_nips, traffic_nips, solar_nips, etth1/2, ettm1/2
+  with horizon ≤30) and long-term forecasting (9 datasets: etth1/2, ettm1/2, traffic_ltsf,
+  electricity_ltsf, exchange_ltsf, illness_ltsf, weather_ltsf with horizons up to
+  720). Training is driven by run.py with per-dataset YAML configs under config/stsf/
+  and config/ltsf/, using PyTorch Lightning. Primary metrics are CRPS (continuous
+  ranked probability score) and NLL (negative log-likelihood). Datasets must be downloaded
+  externally from Google Drive or Baidu Cloud and placed under ./datasets. A typical
+  full run trains 50 epochs per dataset-horizon combination on a single GPU. All evaluation
+  is self-contained within the codebase; no external evaluation toolkit is required.
+
+  '
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- variational_autoencoders
+- probabilistic_forecasting
+- uncertainty_estimation
+- koopman_theory
+- kalman_filtering
+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:
+- crps
+- nll
+- forecast_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval (single dataset)
+  description: 'Run one short-term or long-term K2VAE training + evaluation on a single
+    dataset (e.g., solar_nips or exchange_ltsf) for one prediction length. Expect
+    1-3 hours on a single GPU depending on dataset size.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval (all datasets)
+  description: 'Run the complete K2VAE training and evaluation workflow across all
+    17 datasets and multiple prediction horizons as reported in the paper. This reproduces
+    the full experimental grid (50 epochs per configuration). Expect 12-24+ hours
+    on a single GPU.
+
+    '
+dataset_download:
+  primary_method: manual_download
+  sources:
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/1l0c4H57xYKKQQ5Tm7kd4C8M2nCepky-y?usp=sharing
+    access: public
+    note: Pre-processed time series datasets (ETT, traffic, electricity, exchange,
+      solar, weather, illness). Place contents under ./datasets.
+  - kind: baidu_cloud
+    url: https://pan.baidu.com/s/1eyJvwnPx595nlJzYIzLAJA?pwd=0525
+    access: public
+    note: Mirror of the same datasets on Baidu Cloud (password 0525).
+  notes:
+  - Total dataset size is not documented; expect a few GB after extraction based on
+    the standard ETT/traffic/electricity/weather benchmark datasets.
+  - Download may be slow; the README warns to wait patiently.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - gluonts==0.15.1
+  - lightning
+  notes:
+  - Install via 'pip install .' from the repo root (uses pyproject.toml).
+  - The README recommends optionally running 'pip uninstall -y k2vae' after install
+    to avoid root package conflicts.
+  - See the bundled README and pyproject.toml for the full dependency set.
+risk_flags:
+- manual_dataset_download
+- multi_dataset_grid
+risk_notes:
+- Datasets must be manually downloaded from Google Drive or Baidu Cloud before any
+  run; there is no automated download script.
+- The full experimental grid (17 datasets × multiple horizons × 50 epochs each) is
+  time-intensive; plan for serial or parallel scheduling.
+- No CUDA version is explicitly pinned; PyTorch 2.4.1 supports CUDA 11.8 and 12.x.
+  Verify GPU driver compatibility.
+- The codebase sets torch float32 matmul precision to 'high', which may affect numerical
+  reproducibility across hardware.
+recommended_when: 'Use this benchmark when you want a probabilistic forecasting task
+  with explicit uncertainty calibration metrics (CRPS, NLL), multiple public time-series
+  datasets spanning short- and long-term horizons, and a self-contained training +
+  evaluation loop requiring only a single GPU. Good for exploring Koopman-based latent
+  dynamics or Kalman-filter-inspired uncertainty modeling.
+
+  '
+not_recommended_when: 'Do not use this if you only need deterministic point forecasting
+  without uncertainty quantification, or if you cannot manually download datasets
+  from Google Drive / Baidu Cloud. Also not ideal if you need a benchmark that runs
+  in under an hour for quick iteration.
+
+  '
+paper:
+  title: 'K2VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic
+    Time Series Forecasting'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://arxiv.org/abs/2505.23017
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.053_k2vae.zip
+  archive_type: zip
+  local_dir_name: paper-53-K2VAE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.053_k2vae.zip
+  sha256: 1542a525136c41bdb95128583254d2a8251b57d0a5f161aa92c76a1c0f9ed6e1
+  size_bytes: 10134476
+commercial:
+  annual_fee: null
+display:
+  palette_seed: turquoise-slate-kalman
+  art_style: probabilistic-lab
+  accent_priority: high
+image_path: ../image/053_aisb.t3.053_k2vae.jpg

package/AISB/catalog/aisb.t3.053_k2vae.zh.yaml

@@ -0,0 +1,132 @@
+schema_version: 1
+id: aisb.t3.053_k2vae
+name: 'K2VAE：一种用于概率时序预测的Koopman-Kalman增强变分自编码器'
+version: 0.1.0
+one_line: 在17个公开数据集上训练和评估Koopman-Kalman VAE模型进行概率时间序列预测，衡量CRPS和NLL指标。
+task_description: |
+  本基准测试训练和评估K2VAE，这是一种基于VAE的生成模型，使用KoopmanNet将非线性时间序列线性化为潜在动态系统，
+  并使用KalmanNet通过学习卡尔曼增益来迭代细化预测和建模过程不确定性。该任务涵盖短期预测（8个数据集：
+  exchange_rate_nips、electricity_nips、traffic_nips、solar_nips、etth1/2、ettm1/2，预测长度≤30）
+  和长期预测（9个数据集：etth1/2、ettm1/2、traffic_ltsf、electricity_ltsf、exchange_ltsf、illness_ltsf、
+  weather_ltsf，预测长度最长可达720）。训练由run.py驱动，每个数据集使用config/stsf/和config/ltsf/下的
+  YAML配置文件，采用PyTorch Lightning框架。主要指标为CRPS（连续排名概率得分）和NLL（负对数似然）。
+  数据集需从Google Drive或百度云手动下载并放置于./datasets目录下。典型完整运行需在单GPU上
+  对每个数据集-预测长度组合训练50个epoch。所有评估均在代码库内完成，无需外部评估工具包。
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- variational_autoencoders
+- probabilistic_forecasting
+- uncertainty_estimation
+- koopman_theory
+- kalman_filtering
+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:
+- crps
+- nll
+- forecast_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估（单数据集）
+  description: |
+    在单个数据集（如solar_nips或exchange_ltsf）上运行一次短期或长期K2VAE训练和评估，
+    仅针对一个预测长度。根据数据集大小，预期在单GPU上耗时1-3小时。
+- id: full_train_eval
+  label: 完整训练+评估（全部数据集）
+  description: |
+    在论文中报告的全部17个数据集和多个预测长度上运行完整的K2VAE训练和评估工作流。
+    这将复现完整的实验网格（每个配置50个epoch）。预期在单GPU上耗时12-24小时以上。
+dataset_download:
+  primary_method: manual_download
+  sources:
+  - kind: google_drive
+    url: https://drive.google.com/drive/folders/1l0c4H57xYKKQQ5Tm7kd4C8M2nCepky-y?usp=sharing
+    access: public
+    note: 预处理的时间序列数据集（ETT、交通、电力、汇率、太阳能、天气、疾病）。将内容放置于./datasets目录下。
+  - kind: baidu_cloud
+    url: https://pan.baidu.com/s/1eyJvwnPx595nlJzYIzLAJA?pwd=0525
+    access: public
+    note: 相同数据集在百度云的镜像（密码0525）。
+  notes:
+  - 数据集总大小未记录；根据标准ETT/交通/电力/天气基准数据集，预计解压后为几GB。
+  - 下载可能较慢；README提醒请耐心等待。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - gluonts==0.15.1
+  - lightning
+  notes:
+  - 从仓库根目录执行'pip install .'进行安装（使用pyproject.toml）。
+  - README建议安装后可选执行'pip uninstall -y k2vae'以避免根包冲突。
+  - 完整的依赖列表请参阅附带的README和pyproject.toml。
+risk_flags:
+- manual_dataset_download
+- multi_dataset_grid
+risk_notes:
+  - 数据集必须在任何运行前从Google Drive或百度云手动下载；没有自动下载脚本。
+  - 完整实验网格（17个数据集×多个预测长度×50个epoch）耗时较长；请做好串行或并行调度的规划。
+  - 未明确指定CUDA版本；PyTorch 2.4.1支持CUDA 11.8和12.x。请验证GPU驱动兼容性。
+  - 代码库将torch float32矩阵乘法精度设置为'high'，这可能影响不同硬件间的数值可复现性。
+recommended_when: |
+  当你需要具有明确不确定性校准指标（CRPS、NLL）的概率预测任务、需要涵盖短期和长期预测范围的多个公开时间序列数据集，
+  且需要一个仅需单GPU即可完成的自包含训练和评估循环时，可使用此基准测试。适合探索基于Koopman的潜在动态建模
+  或卡尔曼滤波启发的不确定性建模方法。
+not_recommended_when: |
+  如果你只需要确定性点预测而不需要不确定性量化，或者无法从Google Drive/百度云手动下载数据集，请勿使用此基准测试。
+  如果你需要可在1小时内完成的基准测试以实现快速迭代，此基准测试也不理想。
+paper:
+  title: 'K2VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic
+    Time Series Forecasting'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://arxiv.org/abs/2505.23017
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.053_k2vae.zip
+  archive_type: zip
+  local_dir_name: paper-53-K2VAE
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.053_k2vae.zip
+  sha256: 1542a525136c41bdb95128583254d2a8251b57d0a5f161aa92c76a1c0f9ed6e1
+  size_bytes: 10134476
+commercial:
+  annual_fee: null
+display:
+  palette_seed: turquoise-slate-kalman
+  art_style: probabilistic-lab
+  accent_priority: high
+image_path: ../image/053_aisb.t3.053_k2vae.jpg

package/AISB/catalog/aisb.t3.054_timebase.yaml

@@ -0,0 +1,178 @@
+schema_version: 1
+id: aisb.t3.054_timebase
+name: 'TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting'
+version: 0.1.0
+one_line: 'Train and evaluate TimeBase, an ultra-lightweight (~0.39k parameter) long-term
+  time series forecaster using basis extraction and segment-level prediction, across
+  17+ public benchmarks measuring MSE and MAE.
+
+  '
+task_description: 'This benchmark packages TimeBase, an extremely lightweight long-term
+  time series forecasting (LTSF) model that exploits the approximate low-rank structure
+  of periodic time series. The model segments the input into non-overlapping windows
+  of length P, extracts R orthogonal basis components via a single linear layer, and
+  forecasts future segments via a second linear layer — totaling ~390 parameters.
+  The agent must: (1) download 17 normal-scale benchmark datasets (ETTh1/h2, ETTm1/m2,
+  Weather, Electricity, Traffic, Solar-Energy, Wind, Exchange-Rate, METR-LA, ZafNoo,
+  CzeLan, AQShunyi, AQWan, PM2.5, Temp) from Google Drive and place them in ./dataset/;
+  (2) run per-dataset training scripts via `sh ./scripts/$DATA_NAME.sh`, which invoke
+  `run_longExp.py` → `exp/exp_main.py`; (3) collect avg_mse and avg_mae across horizons
+  (96, 192, 336, 720). Hyperparameter search covers λ_orth ∈ [0.00–0.20] and learning
+  rate ∈ [0.01–0.5]. An optional plug-and-play route applies TimeBase as a complexity
+  reducer for PatchTST via `plug-and-play_for_patchtst/run_all.sh`. Metrics are computed
+  by `utils/metrics.py` using numpy-based MSE/MAE. All evaluation is self-contained;
+  no external evaluation server is required. Optionally, 4 large-scale datasets (CA
+  4.52B, GLA 2.02B, GBA 1.24B, SD 0.38B records) can be obtained from the LargeST
+  repository for extended evaluation.
+
+  '
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- efficient_models
+- forecasting
+- scientific_ml
+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: turnkey
+cost_band: medium
+time_band: 6-24h
+difficulty: medium
+data_access: public
+primary_outputs:
+- avg_mse
+- avg_mae
+- forecast_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run TimeBase training and evaluation on a single normal-scale dataset
+    (e.g., ETTh1) with default hyperparameters. Completes in under 1 hour on a single
+    GPU.
+
+    '
+- id: full_benchmark
+  label: Full 17-Dataset Benchmark
+  description: 'Run all 17 normal-scale dataset scripts sequentially, collecting avg_mse
+    and avg_mae across four forecast horizons (96, 192, 336, 720). Expect 6–24 hours
+    on a single GPU.
+
+    '
+- id: patchtst_plugin
+  label: PatchTST Plug-and-Play
+  description: 'Run the plug-and-play complexity reducer for PatchTST via plug-and-play_for_patchtst/run_all.sh.
+    Demonstrates 77–93% MAC reduction.
+
+    '
+- id: large_scale
+  label: Large-Scale Datasets
+  description: 'Extend evaluation to the 4 billion-record LargeST datasets (CA, GLA,
+    GBA, SD). Requires separate data download from the LargeST GitHub repository and
+    substantially more disk and compute time.
+
+    '
+dataset_download:
+  primary_method: manual
+  sources:
+  - kind: google_drive
+    url: https://drive.google.com/file/d/1ypgCc6iQ2Z8IB_9CY3If_KMRNQKBsI3J/view?usp=sharing
+    access: public
+    note: '17 normal-scale benchmark datasets (ETTh1/h2, ETTm1/m2, Weather, Electricity,
+      Traffic, Solar-Energy, Wind, Exchange-Rate, METR-LA, ZafNoo, CzeLan, AQShunyi,
+      AQWan, PM2.5, Temp). Unzip to ./dataset/.
+
+      '
+  - kind: github
+    url: https://github.com/liuxu77/LargeST
+    access: public
+    note: '4 large-scale datasets (CA, GLA, GBA, SD) with billions of records. Follow
+      LargeST preprocessing instructions. Only needed for the large_scale launch profile.
+
+      '
+  notes:
+  - Normal-scale datasets are modest in size (a few hundred MB total after extraction).
+  - Large-scale datasets require hundreds of GB of disk and are optional.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 1000
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: 3.8+
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - numpy
+  - torch
+  - matplotlib
+  notes:
+  - The model is extremely lightweight (~390 parameters); GPU memory is not the bottleneck.
+  - Recommended disk is 1 TB to accommodate the optional large-scale LargeST datasets.
+  - See the bundled requirements file for the full dependency set.
+  - Python, CUDA, and PyTorch versions are not pinned in the repository; any recent
+    PyTorch with CUDA support should work.
+risk_flags:
+- manual_dataset_download
+- large_dataset_route
+risk_notes:
+- Datasets must be manually downloaded from Google Drive and extracted to ./dataset/
+  before training scripts will run.
+- The 4 large-scale LargeST datasets involve billions of records and require separate
+  preprocessing; attempting them without sufficient disk and RAM will fail.
+- No pinned dependency versions; environment may require manual resolution for older
+  or newer Python/PyTorch combinations.
+recommended_when: 'Use this benchmark when you want an efficiency-focused long-term
+  time series forecasting task on well-established public datasets, or when studying
+  extreme model compression (sub-1k parameters) for LTSF. Good for agents that need
+  a lightweight, self-contained training+evaluation loop with clear numeric metrics
+  (MSE, MAE).
+
+  '
+not_recommended_when: 'Do not use this if you need a multimodal benchmark, a foundation-model-scale
+  task, or a benchmark dominated by large pretrained models. Not suitable if you cannot
+  download data from Google Drive. The large-scale profile is impractical without
+  hundreds of GB of disk.
+
+  '
+paper:
+  title: 'TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting'
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://proceedings.mlr.press/v267/huang25az.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.054_timebase.zip
+  archive_type: zip
+  local_dir_name: paper-54-TimeBase
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.054_timebase.zip
+  sha256: 0871383ee8a2d7dc7f00797e86c0392047d967f3824bf6e76052081150baadce
+  size_bytes: 346725
+commercial:
+  annual_fee: null
+display:
+  palette_seed: ivory-blue-minimal
+  art_style: minimal-forecast
+  accent_priority: medium
+image_path: ../image/054_aisb.t3.054_timebase.jpg