@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.063_tsrag.zh.yaml

@@ -0,0 +1,138 @@
+id: aisb.t3.063_tsrag
+name: 'TS-RAG: 基于检索增强生成的时间序列基础模型是更强的零样本预测器'
+version: 0.1.0
+one_line: 结合时间序列基础模型与知识库的检索增强零样本预测框架，采用自适应检索混合器实现跨领域泛化预测。
+task_description: '该基准测试实现了TS-RAG，一种用于时间序列预测的检索增强生成框架，旨在增强时间序列基础模型（TSFMs）的泛化能力和可解释性。该框架由三个关键组件构成：（1）预训练的时间序列基础模型（TSFM）作为主干预测器；（2）检索器，用于查询专用知识库以查找语义相关的时间序列片段及其未来趋势；（3）可学习的自适应检索混合器（ARM）模块，用于动态融合检索到的模式与TSFM的内部表示。给定输入序列，检索器找到top-k个相似上下文，通过ARM模块进行嵌入和组合，模块自适应分配重要性分数。统一表示被传递到基础模型的输出层，以产生增强的预测结果，无需任务特定的微调。支持chronos-bolt作为主要主干模型。
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 1d+
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- mae
+- mse
+- retrieval_database
+- forecast_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 在已准备的检索数据库上，使用单一主干模型配置运行一个打包的TS-RAG零样本脚本。
+- id: full_rag_eval
+  label: 完整RAG评估
+  description: 运行完整的检索增强零样本预测工作流，包括检索数据库构建、ARM混合和多数据集评估。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - url: https://drive.google.com/drive/folders/12wesXfVwFhdrUY5Kv8yuAWWqN9M77irw
+    type: google_drive
+    description: 预处理后的数据集和预训练模型
+  - url: https://huggingface.co/nkh/TS-RAG-Data
+    type: huggingface
+    description: HuggingFace Hub上的备用镜像
+  notes:
+  - 运行评估脚本前需下载所有数据集和模型检查点
+  - 预期目录结构包括datasets/、retrieval_database/和TS-RAG/checkpoints/
+  - 数据集包括ETT-small、weather、electricity、traffic、exchange_rate和illness
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+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: '3.9'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - chronos-forecasting>=1.5.1
+  - faiss-gpu>=1.7.2
+  - transformers
+  - scikit-learn
+  - pandas
+  - numpy
+  - tqdm
+  notes:
+  - 该工作流依赖于外部检索数据库和预训练的TSFM检查点
+  - 完整的依赖项列表请参阅捆绑的requirements.txt
+  - FAISS-GPU是检索数据库高效相似度搜索所必需的
+  - Chronos管道是默认主干模型所必需的
+risk_flags:
+- external_dependencies
+- large_model_checkpoints
+- retrieval_database_required
+- no_execution_verification
+risk_notes:
+- 预训练的基础模型检查点必须单独下载（chronos-bolt、MOMENT）
+- 检索数据库必须在运行评估前预先构建或生成
+- 打包过程中未执行基准测试；指标值为临时值
+- 外部Google Drive链接可能需要手动下载和组织
+recommended_when: '在以下情况下使用此基准测试：评估检索增强时间序列预测方法、测试基础模型在多样化领域的零样本泛化能力，或基准测试检索上下文模式对预测精度的贡献（无需微调）。
+
+  '
+not_recommended_when: '在以下情况下不要使用此基准测试：无法托管大型检索数据库（FAISS索引）和预训练基础模型检查点；用例仅需要简单的统计基线；或需要在不下载外部资产的情况下获得即时可重复性。
+
+  '
+paper:
+  title: 'TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models
+    are Stronger Zero-Shot Forecaster'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2503.07649
+  authors:
+  - Kanghui Ning
+  - Zijie Pan
+  - Yu Liu
+  - Yushan Jiang
+  - James Y. Zhang
+  - Kashif Rasul
+  - Anderson Schneider
+  - Lintao Ma
+  - Yuriy Nevmyvaka
+  - Dongjin Song
+  affiliation: University of Connecticut (UConn DSIS), Morgan Stanley, Ant Group,
+    TWG Global
+display:
+  palette_seed: steel-jade-retrieval
+  art_style: retrieval-signal
+  accent_priority: high
+image_path: ../image/063_aisb.t3.063_tsrag.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- retrieval_augmented_generation
+- foundation_models
+- zero_shot_learning
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+commercial:
+  annual_fee: null
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.063_tsrag.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.063_tsrag.zip
+  archive_type: zip
+  sha256: 0a4979e3c9db6019ae60f697cd400cdfe7b21f1bb8bc1685d8b1279912aced59
+  size_bytes: 1358610

package/AISB/catalog/aisb.t3.064_pir.yaml

@@ -0,0 +1,221 @@
+id: aisb.t3.064_pir
+name: Improving Time Series Forecasting via Instance-aware Post-hoc Revision
+version: 0.1.0
+one_line: Model-agnostic post-hoc revision framework that identifies and repairs biased
+  time series forecast instances through local and global context-guided revision.
+task_description: 'This packaged benchmark covers PIR (Post-forecasting Identification
+  and Revision), a model-agnostic framework designed to enhance time series forecasting
+  by addressing instance-level variations that mainstream models overlook. The framework
+  operates in two stages: (1) Failure Identification, which estimates per-instance
+  forecasting accuracy to detect potentially biased outputs, and (2) Forecast Revision,
+  which refines identified failures using contextual information from both local and
+  global perspectives. Local revision leverages immediate covariate forecasts and
+  exogenous information within a sliding window; global revision employs retrieval-based
+  strategies to capture long-tail patterns from historical data. The revised forecasts
+  integrate with original outputs via weighted summation, making PIR compatible with
+  arbitrary forecasting backbones including PatchTST, iTransformer, SparseTSF, and
+  TimeMixer. Experiments cover long-term and short-term forecasting settings on established
+  real-world datasets (ETTh1, ETTm1, Weather, Electricity, Traffic).
+
+  '
+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: medium
+data_access: public
+primary_outputs:
+- mse
+- mae
+- revision_outputs
+- forecast_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run one packaged PIR-enhanced forecasting evaluation route on a single
+    backbone and dataset combination.
+- id: full_revision_eval
+  label: Full Revision Eval
+  description: Run the full packaged post-hoc revision workflow across all supported
+    forecasting backbones and datasets.
+dataset_download:
+  primary_method: bundled
+  sources:
+  - name: ETTh1/ETTm1
+    url: https://github.com/thuml/Time-Series-Library
+    note: ETT benchmark from Time-Series-Library
+  - name: Weather
+    url: https://github.com/thuml/Time-Series-Library
+    note: Weather forecasting dataset
+  - name: Electricity
+    url: https://github.com/thuml/Time-Series-Library
+    note: UCI electricity consumption dataset
+  - name: Traffic
+    url: https://github.com/thuml/Time-Series-Library
+    note: PEMS-BAY traffic dataset
+  notes:
+  - Datasets managed via data_provider/data_factory.py and data_loader.py
+  - Data loading follows Time-Series-Library conventions
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.0'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - name: torch
+    min_version: 1.12.0
+  - name: numpy
+    note: For array operations and metrics computation
+  - name: pandas
+    note: For time series data handling
+  notes:
+  - Built on Time-Series-Library (thuml/Time-Series-Library)
+  - See bundled README/requirements for full dependency set
+  - Models implemented in models/ directory (models/PIR.py, models/PatchTST.py, etc.)
+  - Core execution via run.py with exp/exp_long_term_forecasting_pir.py
+  - Metrics computed in utils/metrics.py (MSE, MAE, RMSE, MAPE, MSPE)
+risk_flags:
+- long_tail_patterns
+- distribution_shift
+- instance_variation
+risk_notes:
+- Instance-level variations arise from distribution shifts, missing data, and long-tail
+  patterns
+- Global revision relies on retrieval quality from historical context
+- Weighted integration of original and revised forecasts requires careful tuning per
+  backbone
+recommended_when: 'Use this benchmark when you want to enhance an existing forecasting
+  backbone with post-hoc revision capabilities. Ideal for evaluating model-agnostic
+  forecasting improvements, studying instance-level error patterns, or benchmarking
+  post-processing strategies on established time series datasets.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need a from-scratch foundation-model
+  training task, require a forecasting task without backbones to enhance, or seek
+  methods that operate within the forward pass rather than as post-processing.
+
+  '
+paper:
+  title: Improving Time Series Forecasting via Instance-aware Post-hoc Revision
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.23583
+  authors:
+  - Zhiding Liu
+  - Mingyue Cheng
+  - Guanhao Zhao
+  - Jiqian Yang
+  - Qi Liu
+  - Enhong Chen
+  affiliation: University of Science and Technology of China
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.064_pir.zip
+  archive_type: zip
+  local_dir_name: paper-64-PIR
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.064_pir.zip
+  sha256: 6908afed6d75d39e1d32ceb6504872021a9fc5bf222e115ab78a85fb402716fb
+  size_bytes: 1127481
+display:
+  palette_seed: amber-blue-revise
+  art_style: repair-notebook
+  accent_priority: medium
+image_path: ../image/064_aisb.t3.064_pir.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- post_hoc_revision
+- forecasting
+- instance_level_modeling
+- model_agnostic
+- long_tail_patterns
+- retrieval_augmented
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+commercial:
+  annual_fee: null
+code_anchors:
+  executable:
+  - path: run.py
+    role: Entry point for training and evaluation
+  - path: exp/exp_long_term_forecasting_pir.py
+    role: PIR experiment class implementing training/evaluation loop
+  - path: utils/metrics.py
+    role: Metric functions (MSE, MAE, RMSE, MAPE, MSPE)
+  data:
+  - path: data_provider/data_factory.py
+    role: Data loader factory
+  - path: data_provider/data_loader.py
+    role: Time series data loading
+  models:
+  - path: models/PIR.py
+    role: PIR model implementation
+  - path: models/PatchTST.py
+    role: PatchTST backbone
+  - path: models/iTransformer.py
+    role: iTransformer backbone
+  - path: models/SparseTSF.py
+    role: SparseTSF backbone
+  - path: models/TimeMixer.py
+    role: TimeMixer backbone
+  scripts:
+  - path: scripts/
+    role: Shell scripts for running backbone and PIR-enhanced experiments per dataset
+metric_contract:
+  mse:
+    origin_path: utils/metrics.py
+    source_ref: MSE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+  mae:
+    origin_path: utils/metrics.py
+    source_ref: MAE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+  rmse:
+    origin_path: utils/metrics.py
+    source_ref: RMSE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+notes:
+- PIR operates as a post-processing plugin, not requiring modifications to the base
+  forecaster
+- The framework handles both long-term and short-term forecasting scenarios
+- Integration with Time-Series-Library ensures compatibility with standard benchmark
+  protocols
+- Static code audit confirms executable anchors for all staged metrics
+- No benchmark execution was performed during packaging pass; runtime validation recommended

package/AISB/catalog/aisb.t3.064_pir.zh.yaml

@@ -0,0 +1,197 @@
+id: aisb.t3.064_pir
+name: 通过实例感知的后处理修正提升时间序列预测
+version: 0.1.0
+one_line: 一种模型无关的后处理框架，通过局部和全局上下文引导的修正，识别并修复有偏的时间序列预测实例。
+task_description: '本基准测试包涵盖 PIR（预测后识别与修正，Post-forecasting Identification and Revision），这是一个模型无关的框架，旨在通过处理主流模型忽视的实例级变异来增强时间序列预测。该框架分两个阶段运行：（1）失败识别，通过估计每个实例的预测准确性来检测可能有偏的输出；（2）预测修正，利用局部和全局视角的上下文信息来优化已识别的失败。局部修正利用滑动窗口内的即时协变量预测和外生信息；全局修正采用基于检索的策略从历史数据中捕捉长尾模式。修正后的预测通过加权求和与原始输出整合，使 PIR 可与任意预测骨干网络兼容，包括 PatchTST、iTransformer、SparseTSF 和 TimeMixer。实验涵盖在已有真实数据集（ETTh1、ETTm1、Weather、Electricity、Traffic）上的长期和短期预测设置。
+
+  '
+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: medium
+data_access: public
+primary_outputs:
+- mse
+- mae
+- revision_outputs
+- forecast_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 在单一骨干网络和数据集组合上运行一个打包的 PIR 增强预测评估流程。
+- id: full_revision_eval
+  label: 完整修正评估
+  description: 在所有支持的预测骨干网络和数据集上运行完整的后处理修正工作流。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - name: ETTh1/ETTm1
+    url: https://github.com/thuml/Time-Series-Library
+    note: 来自 Time-Series-Library 的 ETT 基准数据集
+  - name: Weather
+    url: https://github.com/thuml/Time-Series-Library
+    note: 天气预测数据集
+  - name: Electricity
+    url: https://github.com/thuml/Time-Series-Library
+    note: UCI 电力消耗数据集
+  - name: Traffic
+    url: https://github.com/thuml/Time-Series-Library
+    note: PEMS-BAY 交通数据集
+  notes:
+  - 数据集通过 data_provider/data_factory.py 和 data_loader.py 管理
+  - 数据加载遵循 Time-Series-Library 规范
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.0'
+  pytorch: 1.12.0
+  flash_attn: null
+  key_packages:
+  - name: torch
+    min_version: 1.12.0
+  - name: numpy
+    note: 用于数组操作和指标计算
+  - name: pandas
+    note: 用于时间序列数据处理
+  notes:
+  - 基于 Time-Series-Library 构建（thuml/Time-Series-Library）
+  - 参见打包的 README/requirements 以获取完整依赖项
+  - 模型在 models/ 目录中实现（models/PIR.py、models/PatchTST.py 等）
+  - 核心执行通过 run.py 和 exp/exp_long_term_forecasting_pir.py
+  - 指标在 utils/metrics.py 中计算（MSE、MAE、RMSE、MAPE、MSPE）
+risk_flags:
+- long_tail_patterns
+- distribution_shift
+- instance_variation
+risk_notes:
+- 实例级变异源于分布偏移、数据缺失和长尾模式
+- 全局修正依赖于从历史上下文中检索的质量
+- 原始预测和修正预测的加权整合需要对每个骨干网络进行仔细调优
+recommended_when: '当您希望使用后处理修正能力增强现有预测骨干网络时使用此基准测试。适用于评估模型无关的预测改进、研究实例级误差模式，或在已有时间序列数据集上基准化后处理策略。
+
+  '
+not_recommended_when: '如果您需要从头开始的基础模型训练任务、需要不带骨干网络增强的预测任务，或寻求在前向传播内而非后处理中运行的方法，请勿使用此基准测试。
+
+  '
+paper:
+  title: Improving Time Series Forecasting via Instance-aware Post-hoc Revision
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.23583
+  authors:
+  - Zhiding Liu
+  - Mingyue Cheng
+  - Guanhao Zhao
+  - Jiqian Yang
+  - Qi Liu
+  - Enhong Chen
+  affiliation: University of Science and Technology of China
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.064_pir.zip
+  archive_type: zip
+  local_dir_name: paper-64-PIR
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.064_pir.zip
+  sha256: 6908afed6d75d39e1d32ceb6504872021a9fc5bf222e115ab78a85fb402716fb
+  size_bytes: 1127481
+display:
+  palette_seed: amber-blue-revise
+  art_style: repair-notebook
+  accent_priority: medium
+image_path: ../image/064_aisb.t3.064_pir.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- post_hoc_revision
+- forecasting
+- instance_level_modeling
+- model_agnostic
+- long_tail_patterns
+- retrieval_augmented
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+commercial:
+  annual_fee: null
+code_anchors:
+  executable:
+  - path: run.py
+    role: Entry point for training and evaluation
+  - path: exp/exp_long_term_forecasting_pir.py
+    role: PIR experiment class implementing training/evaluation loop
+  - path: utils/metrics.py
+    role: Metric functions (MSE, MAE, RMSE, MAPE, MSPE)
+  data:
+  - path: data_provider/data_factory.py
+    role: Data loader factory
+  - path: data_provider/data_loader.py
+    role: Time series data loading
+  models:
+  - path: models/PIR.py
+    role: PIR model implementation
+  - path: models/PatchTST.py
+    role: PatchTST backbone
+  - path: models/iTransformer.py
+    role: iTransformer backbone
+  - path: models/SparseTSF.py
+    role: SparseTSF backbone
+  - path: models/TimeMixer.py
+    role: TimeMixer backbone
+  scripts:
+  - path: scripts/
+    role: Shell scripts for running backbone and PIR-enhanced experiments per dataset
+metric_contract:
+  mse:
+    origin_path: utils/metrics.py
+    source_ref: MSE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+  mae:
+    origin_path: utils/metrics.py
+    source_ref: MAE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+  rmse:
+    origin_path: utils/metrics.py
+    source_ref: RMSE function
+    evaluation_protocol:
+      code_paths:
+      - exp/exp_long_term_forecasting_pir.py
+      - utils/metrics.py
+    status: code_backed
+notes:
+- PIR 作为后处理插件运行，无需修改基础预测器
+- 该框架同时处理长期和短期预测场景
+- 与 Time-Series-Library 的集成确保了与标准基准测试协议的兼容性
+- 静态代码审计确认所有分阶段指标的可执行锚点
+- 打包过程中未执行基准测试；建议进行运行时验证