@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.025_deeper.yaml

@@ -0,0 +1,143 @@
+id: aisb.t3.025_deeper
+name: 'DEEPER: Directed Persona Refinement for Dynamic Persona Modeling'
+version: 0.1.0
+one_line: Novel refinement-based dynamic persona modeling using offline reinforcement
+  learning and discrepancy-driven update direction search across 10 domains.
+task_description: 'This packaged benchmark covers dynamic persona modeling with a
+  novel refinement-based paradigm that addresses fundamental limitations of traditional
+  regeneration (replacing personas) and extension (incrementally appending behaviors)
+  approaches. DEEPER leverages prediction-behavior discrepancies to guide update directions
+  in a structured, reward-driven way, enabling continual persona optimization over
+  time. The method employs a tri-objective reward design balancing Previous Preservation,
+  Current Reflection, and Future Advancement, trained via two-stage iterative offline
+  RL with DPO fine-tuning. Evaluated on 4,800+ users across 10 domains including 4
+  unseen domains for cross-domain generalization testing. Primary evaluation metric
+  is MAE on future behavior prediction across four refinement rounds.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- mae_round4
+- persona_refinement_model
+- evaluation_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged evaluation pass on the DEEPER persona modeling stack
+    using eval_round4.py. Evaluates MAE on Round 4 (predicting window 5 ratings using
+    persona S4 after 4 DEEPER updates). Uses meta-llama/llama-3.3-70b-instruct via
+    OpenRouter API with configurable concurrency.
+
+    '
+- id: full_refinement
+  label: Full Refinement
+  description: 'Run the complete persona refinement workflow including two-stage offline
+    training (iteration 1 and iteration 2) via LLaMA-Factory, persona updates across
+    all rounds, and downstream evaluation with MAE reporting per domain and round.
+
+    '
+dataset_download:
+  primary_method: huggingface
+  sources:
+  - https://huggingface.co/datasets/deeper-team/DEEPER_preprocess_data
+  - https://huggingface.co/datasets/deeper-team/DEEPER_user_context_data
+  - https://huggingface.co/datasets/deeper-team/DEEPER_train_data
+  - https://huggingface.co/deeper-team/DEEPER-llama-8B
+  notes:
+  - Four dataset components: Preprocessed Data, User Context Data, DEEPER Training
+      Data, Evolving Personas
+  - Preprocessed data spans 10 domains with chronological user rating histories (50+
+    entries per user)
+  - User context data organized by iteration (iteration_1 through iteration_4 for
+    eval, iteration_1 through iteration_2 for train)
+  - Training data constructed via self-sampling and iterative optimization over two
+    training iterations
+credential_requirements:
+  mode: api_key
+  items:
+  - openrouter_api_key
+  notes:
+  - OpenRouter API key required for LLM-based persona generation and behavior prediction
+    during evaluation
+  - API calls use meta-llama/llama-3.3-70b-instruct model
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 48
+    gpu_type: NVIDIA A100 or equivalent with high VRAM
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 2
+    gpu_vram_gb: 80
+    gpu_type: Multi-GPU setup for parallel training and inference
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - transformers==4.45.0
+  - vllm==0.7.3
+  - openai>=1.0.0
+  - llama-factory @ git+https://github.com/hiyouga/LLaMA-Factory.git
+  notes:
+  - LLaMA-Factory required for training pipeline (install via git clone and pip install
+    -e ".[torch,metrics]")
+  - Conda environment strongly recommended for dependency isolation
+  - See bundled README/requirements.txt for complete dependency set
+risk_flags:
+- multi_gpu_training
+- llm_api_dependency
+- rate_limiting
+- long_runtime
+risk_notes:
+- Multi-GPU training orchestration increases failure complexity; checkpointing recommended
+- Evaluation relies on OpenRouter API calls; rate limits may cause timeouts at MAX_WORKERS=8
+- Full refinement workflow spans multiple training iterations and 4 evaluation rounds
+- Train iterations use DPO fine-tuning which is sensitive to hyperparameters
+recommended_when: 'Use this benchmark when you need a personalization task that combines
+  behavior modeling, preference optimization, and cross-domain generalization testing.
+  Ideal for evaluating offline RL methods on dynamic user modeling scenarios requiring
+  iterative persona refinement with explicit direction search capabilities.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you cannot host or access 8B-class
+  LLMs, if you lack API infrastructure for OpenRouter calls, or if you need a quick
+  baseline task without RL components. Not suitable for environments without GPU capability
+  or with strict compute budgets below the minimum resource requirements.
+
+  '
+paper:
+  title: 'DEEPER: Directed Persona Refinement for Dynamic Persona Modeling'
+  venue: ACL 2025
+  year: 2025
+  url: https://arxiv.org/abs/2502.11078
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.025_deeper.zip
+  archive_type: zip
+  local_dir_name: paper-25-DEEPER
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.025_deeper.zip
+  sha256: 1c2f13d7dc01d3ff88b03778010e1d29b20d5bb4b687e9953a0b0b0977b0057e
+  size_bytes: 3096009
+display:
+  palette_seed: plum-sand-persona
+  art_style: human-centered
+  accent_priority: high
+image_path: ../image/025_aisb.t3.025_deeper.jpg

package/AISB/catalog/aisb.t3.025_deeper.zh.yaml

@@ -0,0 +1,116 @@
+id: aisb.t3.025_deeper
+name: 'DEEPER：用于动态人格建模的导向性人格优化方法'
+version: 0.1.0
+one_line: 基于优化的新型动态人格建模方法，采用离线强化学习和差异驱动的更新方向搜索，涵盖10个领域。
+task_description: '该基准测试包涵盖动态人格建模，采用一种基于优化的新范式，解决了传统再生（替换人格）和扩展（增量追加行为）方法的根本性局限。DEEPER利用预测-行为差异以结构化、奖励驱动的方式引导更新方向，实现人格的持续优化。该方法采用三目标奖励设计，平衡"历史保持"、"当前反映"和"未来推进"，通过两阶段迭代离线强化学习和DPO微调进行训练。在4800+用户和10个领域（包括4个未见领域用于跨域泛化测试）上进行评估。主要评估指标是在四轮优化过程中对未来行为预测的MAE。
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- mae_round4
+- persona_refinement_model
+- evaluation_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: '使用eval_round4.py对DEEPER人格建模栈运行一次打包评估。评估第4轮MAE（使用经过4次DEEPER更新后的人格S4预测第5窗口评分）。通过OpenRouter API使用meta-llama/llama-3.3-70b-instruct，支持可配置的并发数。
+
+    '
+- id: full_refinement
+  label: 完整优化流程
+  description: '运行完整的人格优化工作流，包括通过LLaMA-Factory进行两阶段离线训练（迭代1和迭代2）、所有轮次的人格更新，以及包含各领域和各轮次MAE报告的下游评估。
+
+    '
+dataset_download:
+  primary_method: huggingface
+  sources:
+  - https://huggingface.co/datasets/deeper-team/DEEPER_preprocess_data
+  - https://huggingface.co/datasets/deeper-team/DEEPER_user_context_data
+  - https://huggingface.co/datasets/deeper-team/DEEPER_train_data
+  - https://huggingface.co/deeper-team/DEEPER-llama-8B
+  notes:
+  - 四个数据集组件：预处理数据、用户上下文数据、DEEPER训练数据、演化人格
+  - 预处理数据涵盖10个领域，包含按时间顺序的用户评分历史（每用户50+条记录）
+  - 用户上下文数据按迭代组织（评估用iteration_1到iteration_4，训练用iteration_1到iteration_2）
+  - 训练数据通过自采样和两轮迭代优化构建
+credential_requirements:
+  mode: api_key
+  items:
+  - openrouter_api_key
+  notes:
+  - 需要OpenRouter API密钥用于评估期间基于LLM的人格生成和行为预测
+  - API调用使用meta-llama/llama-3.3-70b-instruct模型
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 48
+    gpu_type: NVIDIA A100或等效高显存GPU
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 2
+    gpu_vram_gb: 80
+    gpu_type: 多GPU配置用于并行训练和推理
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - transformers==4.45.0
+  - vllm==0.7.3
+  - openai>=1.0.0
+  - llama-factory @ git+https://github.com/hiyouga/LLaMA-Factory.git
+  notes:
+  - LLaMA-Factory是训练流程必需的（通过git clone和pip install -e ".[torch,metrics]"安装）
+  - 强烈建议使用Conda环境进行依赖隔离
+  - 参见附带的README/requirements.txt获取完整的依赖列表
+risk_flags:
+- multi_gpu_training
+- llm_api_dependency
+- rate_limiting
+- long_runtime
+risk_notes:
+- 多GPU训练编排增加故障复杂性；建议使用检查点
+- 评估依赖OpenRouter API调用；在MAX_WORKERS=8时可能因速率限制导致超时
+- 完整优化工作流跨越多轮训练迭代和4轮评估
+- 训练迭代使用DPO微调，对超参数敏感
+recommended_when: '当您需要一个结合行为建模、偏好优化和跨域泛化测试的个性化任务时使用此基准。特别适合评估离线强化学习方法在需要迭代人格优化和显式方向搜索能力的动态用户建模场景中的表现。
+
+  '
+not_recommended_when: '如果您无法托管或访问8B级LLM、缺乏OpenRouter调用的API基础设施，或需要一个不含RL组件的快速基线任务，请勿使用此基准。不适合没有GPU能力或计算预算低于最低资源要求的环境。
+
+  '
+paper:
+  title: 'DEEPER: Directed Persona Refinement for Dynamic Persona Modeling'
+  venue: ACL 2025
+  year: 2025
+  url: https://arxiv.org/abs/2502.11078
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.025_deeper.zip
+  archive_type: zip
+  local_dir_name: paper-25-DEEPER
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.025_deeper.zip
+  sha256: 1c2f13d7dc01d3ff88b03778010e1d29b20d5bb4b687e9953a0b0b0977b0057e
+  size_bytes: 3096009
+display:
+  palette_seed: plum-sand-persona
+  art_style: human-centered
+  accent_priority: high
+image_path: ../image/025_aisb.t3.025_deeper.jpg

package/AISB/catalog/aisb.t3.026_gartkg.yaml

@@ -0,0 +1,195 @@
+schema_version: 1
+id: aisb.t3.026_gartkg
+name: DGAR – Deep Generative Adaptive Replay for Temporal Knowledge Graph Reasoning
+version: 0.1.0
+one_line: 'Continual temporal knowledge-graph reasoning benchmark: train a diffusion-guided
+  adaptive replay model (DGAR) on ICEWS14s-style TKG snapshots and measure link-prediction
+  accuracy (MRR, Hits@1/3/10) under sequential task arrival with catastrophic-forgetting
+  mitigation.
+
+  '
+task_description: 'This benchmark packages the DGAR method for continual Temporal
+  Knowledge Graph Reasoning (TKGR). The agent must (1) build Historical Context Prompts
+  from past KG snapshots, (2) pre-train a diffusion model that generates historical
+  entity distributions, (3) run continual learning over a stream of TKG snapshots
+  using adaptive replay to mitigate catastrophic forgetting, and (4) evaluate link-prediction
+  quality (MRR, Hits@1, Hits@3, Hits@10) on held-out test splits at each time step.
+  The core entry point is `src/main.py --gpu 0 --dataset ICEWS14s`. The model combines
+  RE-GCN graph reasoning with a Transformer-based DDPM for entity-distribution generation
+  and a layer-by-layer deep adaptive replay mechanism. Hyperparameter ranges are defined
+  in `src/hyperparameter_range.py`. The primary metric (`accuracy` / MRR) is code-backed
+  via `diffusion/difffu_21.py`. No benchmark execution was performed during the packaging
+  pass; the agent must execute and validate all metrics at runtime.
+
+  '
+capability_tags:
+- research_code_optimization
+- temporal_kg_reasoning
+- graph_learning
+- continual_learning
+- generative_replay
+- diffusion_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_not_verified
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- mrr_filter
+- hits_at_1
+- hits_at_3
+- hits_at_10
+- temporal_kg_checkpoint
+- reasoning_report
+launch_profiles:
+- id: quick_check
+  label: Quick Check
+  description: 'Run a single dataset (e.g. ICEWS14s) through `src/main.py` with default
+    hyperparameters for one pass over the continual task stream. Verifies that the
+    training loop, diffusion generation, and evaluation pipeline execute end-to-end.
+    Expected wall-time: 1–3 hours on a single GPU.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval
+  description: 'Run the full continual training and evaluation workflow including
+    hyperparameter sweeps over n_hidden, n_layers, and dropout (see `src/hyperparameter_range.py`).
+    Covers all TKG snapshot tasks with adaptive replay, diffusion-enhanced generation,
+    and deep adaptive replay. Expected wall-time: 6–24 hours depending on dataset
+    size and GPU.
+
+    '
+- id: hp_sweep
+  label: Hyperparameter Sweep
+  description: 'Systematic grid search over the ranges in `src/hyperparameter_range.py`:
+    n_hidden ∈ {100,200,300,400}, n_layers ∈ {1,2}, dropout ∈ {0.2,0.4}, n_bases=100.
+    Produces a comparison table of MRR and Hits@k across configurations.
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources:
+  - kind: local
+    url: null
+    access: public
+    note: 'Dataset files (e.g. ICEWS14s) are expected to be bundled in the snapshot
+      or loadable via the code''s built-in data-reading utilities (see `rgcn/knowledge_graph.py`).
+      The README indicates running with `--dataset ICEWS14s`.
+
+      '
+  notes:
+  - Dataset size is modest (tens of MB for ICEWS14s). No large external download is
+    required.
+  - If datasets are not bundled, check the original ICEWS14s public sources.
+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:
+  - dgl
+  - fitlog
+  - umap-learn
+  - scikit-learn
+  - scipy
+  - tqdm
+  - matplotlib
+  notes:
+  - See bundled requirements for exact pinned versions.
+  - Code imports dgl, fitlog, umap, sklearn, scipy – all must be installed.
+  - The diffusion module uses a Transformer-based architecture; no external pretrained
+    weights are referenced.
+risk_flags:
+- unverified_execution
+- implicit_dataset_assumption
+- fitlog_dependency
+risk_notes:
+- 'No benchmark execution was performed during the packaging pass. All metric values
+  must be produced and validated by the agent at runtime.
+
+  '
+- 'The README entry point has a formatting artifact (`--dataset ICEWS14s# DGAR`);
+  the actual flag is likely `--dataset ICEWS14s`. The agent should verify the correct
+  dataset name.
+
+  '
+- 'fitlog is used for experiment tracking; if not installed or configured it may cause
+  import errors.
+
+  '
+- 'Code depends on dgl (Deep Graph Library) which has CUDA-version-specific builds.
+  Ensure the dgl version matches the installed CUDA toolkit.
+
+  '
+- 'The diffusion model pre-training and TKGR training are interleaved in the continual
+  learning loop; GPU memory may spike during the diffusion generation + gradient guidance
+  step (Eq. 8–10 in paper).
+
+  '
+recommended_when: 'Use this benchmark when you want a graph-learning task that combines
+  continual learning with generative replay and temporal knowledge graph reasoning.
+  Suitable for evaluating agents that must navigate multi-component ML pipelines (GNN
+  + diffusion + replay) with hyperparameter optimization under a continual-learning
+  protocol.
+
+  '
+not_recommended_when: 'Not suitable if you need a lightweight text-only task, cannot
+  provide ≥16 GB GPU VRAM for the combined graph + diffusion components, or require
+  a benchmark with pre-verified baseline metric values. Also not ideal if you need
+  multi-GPU distributed training support (code appears single-GPU only).
+
+  '
+paper:
+  title: A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge
+    Graph Reasoning
+  authors:
+  - Zhiyu Zhang
+  - Wei Chen
+  - Youfang Lin
+  - Huaiyu Wan
+  venue: ACL 2025
+  year: 2025
+  url: https://aclanthology.org/2025.acl-long.537/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.026_gartkg.zip
+  archive_type: zip
+  local_dir_name: paper-26-GARTKG
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.026_gartkg.zip
+  sha256: 010bac738fef854266a59273862289b34ebafc7890602fd29f73613783e594f2
+  size_bytes: 81352
+commercial:
+  annual_fee: null
+display:
+  palette_seed: forest-copper-temporal
+  art_style: graph-science
+  accent_priority: high
+image_path: ../image/026_aisb.t3.026_gartkg.jpg

package/AISB/catalog/aisb.t3.026_gartkg.zh.yaml

@@ -0,0 +1,127 @@
+schema_version: 1
+id: aisb.t3.026_gartkg
+name: DGAR – 用于时序知识图谱推理的深度生成式自适应回放
+version: 0.1.0
+one_line: 持续性时序知识图谱推理基准：通过自适应回放训练扩散引导模型(DGAR)，在ICEWS14s风格的时间知识图谱快照上完成链接预测任务，在顺序任务到达场景下评估MRR和Hits@1/3/10指标，同时实现灾难性遗忘缓解。
+task_description: 本基准测试封装了DGAR方法用于持续性时序知识图谱推理(TKGR)。智能体需要完成以下工作：(1)从历史知识图谱快照构建上下文提示，(2)预训练扩散模型以生成历史实体分布，(3)在时间知识图谱快照流上运行持续学习，使用自适应回放机制缓解灾难性遗忘，(4)在每个时间步骤的保留测试集上评估链接预测质量(MRR, Hits@1, Hits@3, Hits@10)。核心入口为 `src/main.py --gpu 0 --dataset ICEWS14s`。模型结合了RE-GCN图推理、基于Transformer的DDPM用于实体分布生成，以及逐层深度自适应回放机制。超参数范围定义在 `src/hyperparameter_range.py` 中。主指标(`accuracy`/MRR)通过 `diffusion/difffu_21.py` 实现代码验证。打包过程中未执行基准测试；智能体必须在运行时执行并验证所有指标。
+capability_tags:
+- research_code_optimization
+- temporal_kg_reasoning
+- graph_learning
+- continual_learning
+- generative_replay
+- diffusion_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_not_verified
+support_level: advanced
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- mrr_filter
+- hits_at_1
+- hits_at_3
+- hits_at_10
+- temporal_kg_checkpoint
+- reasoning_report
+launch_profiles:
+- id: quick_check
+  label: 快速检查
+  description: 使用默认超参数在单个数据集(如ICEWS14s)上运行 `src/main.py`，对持续任务流执行一次完整遍历。验证训练循环、扩散生成和评估管道的端到端执行。预计耗时：单GPU 1-3小时。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: 运行完整的持续训练和评估工作流，包括对n_hidden、n_layers和dropout的超参数搜索(见 `src/hyperparameter_range.py`)。覆盖所有时间知识图谱快照任务，包含自适应回放、扩散增强生成和深度自适应回放。预计耗时：6-24小时，取决于数据集大小和GPU配置。
+- id: hp_sweep
+  label: 超参数搜索
+  description: 对 `src/hyperparameter_range.py` 中的范围进行系统性网格搜索：n_hidden ∈ {100,200,300,400}, n_layers ∈ {1,2}, dropout ∈ {0.2,0.4}, n_bases=100。生成不同配置下MRR和Hits@k的对比表。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - kind: local
+    url: null
+    access: public
+    note: 数据集文件(如ICEWS14s)预期随快照打包提供，或可通过代码内置的数据读取工具加载(见 `rgcn/knowledge_graph.py`)。README指示使用 `--dataset ICEWS14s` 运行。
+  notes:
+  - 数据集规模较小(ICEWS14s仅数十MB)。无需下载大型外部文件。
+  - 如数据集未打包，请查阅原始ICEWS14s公开数据源。
+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:
+  - dgl
+  - fitlog
+  - umap-learn
+  - scikit-learn
+  - scipy
+  - tqdm
+  - matplotlib
+  notes:
+  - 详见打包的requirements文件以获取精确的版本锁定。
+  - 代码依赖dgl、fitlog、umap、sklearn、scipy——全部必须安装。
+  - 扩散模块使用基于Transformer的架构；不引用外部预训练权重。
+risk_flags:
+- unverified_execution
+- implicit_dataset_assumption
+- fitlog_dependency
+risk_notes:
+- 打包过程中未执行基准测试。所有指标值必须由智能体在运行时生成并验证。
+- README入口点存在格式问题(`--dataset ICEWS14s# DGAR`)；实际参数可能是 `--dataset ICEWS14s`。智能体应验证正确的数据集名称。
+- fitlog用于实验追踪；如果未安装或未配置可能导致导入错误。
+- 代码依赖dgl(深度图库)，它有针对CUDA版本特定构建的版本。请确保dgl版本与已安装的CUDA工具包匹配。
+- 扩散模型预训练和知识图谱推理训练在持续学习循环中交错进行；在扩散生成+梯度引导步骤(论文中的公式8-10)期间GPU显存可能达到峰值。
+recommended_when: 当你需要结合持续学习和生成式回放的图学习任务，用于时间知识图谱推理时使用此基准。适用于评估需要处理多组件ML流程(GNN + 扩散 + 回放)且在持续学习协议下进行超参数优化的智能体。
+not_recommended_when: 如果你需要轻量级纯文本任务、无法提供≥16GB GPU显存来支持图+扩散组件组合，或需要一个预验证基准指标值的基准测试，则不适合使用。同样，如果你需要多GPU分布式训练支持(代码似乎仅支持单GPU)，也不理想。
+paper:
+  title: A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning
+  authors:
+  - Zhiyu Zhang
+  - Wei Chen
+  - Youfang Lin
+  - Huaiyu Wan
+  venue: ACL 2025
+  year: 2025
+  url: https://aclanthology.org/2025.acl-long.537/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.026_gartkg.zip
+  archive_type: zip
+  local_dir_name: paper-26-GARTKG
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.026_gartkg.zip
+  sha256: 010bac738fef854266a59273862289b34ebafc7890602fd29f73613783e594f2
+  size_bytes: 81352
+commercial:
+  annual_fee: null
+display:
+  palette_seed: forest-copper-temporal
+  art_style: graph-science
+  accent_priority: high
+image_path: ../image/026_aisb.t3.026_gartkg.jpg