@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.058_mindglitch.yaml

@@ -0,0 +1,171 @@
+id: aisb.t3.058_mindglitch
+name: 'Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in Subject-Driven
+  Generation'
+version: 0.1.0
+one_line: Disentangle visual and semantic features from diffusion backbones to quantify
+  and localize inconsistencies in subject-driven image generation.
+task_description: 'This benchmark implements the Mind-the-Glitch framework from NeurIPS
+  2025, which disentangles visual and semantic features from pre-trained diffusion
+  model backbones to enable visual correspondence matching. The core contribution
+  is the Visual Semantic Matching (VSM) metric, which quantifies the degree of visual
+  inconsistency between a reference subject and generated images while also supporting
+  spatial localization of inconsistent regions. Unlike global feature-based metrics
+  (CLIP, DINO) or VLM-based approaches, VSM captures fine-grained appearance-level
+  variations independent of pose, scale, or context. The package includes pre-trained
+  MTG model weights, inference code, and a playground notebook demonstrating feature
+  extraction, correspondence computation, and inconsistency scoring. Training code,
+  training dataset generation pipeline, and full benchmark evaluation code are not
+  yet released.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- spearman_correlation
+- pearson_correlation
+- vsm_scores
+- mtg_inference_scores
+launch_profiles:
+- id: playground
+  label: Playground
+  description: 'Run the interactive playground notebook to load the pre-trained MTG
+    model, perform inference on sample image pairs, visualize disentangled semantic
+    and visual features, and compute VSM scores for inconsistency detection.
+
+    '
+- id: inference_eval
+  label: Inference Eval
+  description: 'Run the packaged evaluation scripts (eval_run_iter10_tta.py, eval_run_iter12_weighted_tta.py)
+    to compute Spearman and Pearson correlations between VSM metric and oracle judgments
+    on the test dataset. Note that full training pipeline and benchmark evaluation
+    code are not yet released in this snapshot.
+
+    '
+dataset_download:
+  primary_method: git_huggingface
+  sources:
+  - type: github
+    url: https://github.com/abdo-eldesokey/mind-the-glitch
+    note: Clone with submodules; requires Grounded-Segment-Anything submodule setup
+  - type: huggingface
+    url: https://huggingface.co/mind-the-glitch
+    note: Pre-trained model weights and safetensors
+  notes:
+  - Clone recursively: git clone --recursive
+  - Initialize submodules after clone if needed
+  - Set HF_HOME environment variable for model cache
+  - HF token may be required for some model downloads
+credential_requirements:
+  mode: optional_token
+  items:
+  - huggingface_token
+  notes:
+  - HF_TOKEN may be needed for gated model access; an anonymous token is used in eval
+    scripts
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    cuda_capability: '8.0'
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    cuda_capability: '8.6'
+environment:
+  python: '3.11'
+  cuda: '12.4'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - torch==2.5.1
+  - torchvision==0.20.1
+  - omegaconf
+  - hydra-core
+  - safetensors
+  - scipy
+  - Pillow
+  - numpy
+  - scikit-image
+  - GroundingDINO
+  - segment-anything
+  notes:
+  - Grounded-Segment-Anything must be installed separately after cloning
+  - See bundled requirements.txt for complete dependency list
+  - HF endpoint configured to mirror for faster downloads in eval scripts
+risk_flags:
+- inference_only
+- no_training_release
+- no_benchmark_eval_release
+risk_notes:
+- Current snapshot provides inference-only functionality
+- Training code and training dataset generation pipeline are not released
+- Full benchmark evaluation framework is not yet included
+- Static code audit confirms executable anchors exist for staged metrics
+- No runtime execution was performed during this packaging pass
+recommended_when: 'Use this benchmark when you need to evaluate subject consistency
+  in subject-driven image generation, compare VSM against existing metrics (CLIP,
+  DINO, LPIPS), or leverage visual correspondence for localizing inconsistent regions
+  within generated subjects.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you require a complete training
+  pipeline, need to train or fine-tune the MTG model, or require the full benchmark
+  evaluation framework with standardized test sets and oracle comparisons. The current
+  release is inference-oriented only.
+
+  '
+paper:
+  title: 'Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in
+    Subject-Driven Generation'
+  authors:
+  - Abdelrahman Eldesokey
+  - Aleksandar Cvejic
+  - Bernard Ghanem
+  - Peter Wonka
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2509.21989
+  project_url: https://abdo-eldesokey.github.io/mind-the-glitch/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.058_mindglitch.zip
+  archive_type: zip
+  local_dir_name: paper-58-MindGlitch
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.058_mindglitch.zip
+  sha256: 26309a90f48ab6a8b73a083a10ef1e7c2d3aed96f48a9bbf1bcf1cc2cd807b89
+  size_bytes: 3505243
+display:
+  palette_seed: mauve-graphite-correspond
+  art_style: visual-consistency
+  accent_priority: high
+image_path: ../image/058_aisb.t3.058_mindglitch.jpg
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+capability_tags:
+- computer_vision
+- diffusion_models
+- visual_correspondence
+- consistency_detection
+- subject_driven_generation
+- feature_disentanglement
+- vsm_metric
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.058_mindglitch.zh.yaml

@@ -0,0 +1,145 @@
+id: aisb.t3.058_mindglitch
+name: 'Mind-the-Glitch：视觉一致性检测中的不一致区域发现'
+version: 0.1.0
+one_line: 从扩散骨干网络中解耦视觉和语义特征，以量化和定位主体驱动图像生成中的不一致区域。
+task_description: '本基准实现了NeurIPS 2025的Mind-the-Glitch框架，该框架从预训练扩散模型骨干网络中解耦视觉和语义特征，以实现视觉对应匹配。核心贡献是视觉语义匹配（VSM）指标，它能量化参考主体与生成图像之间的视觉不一致程度，同时支持不一致区域的空间定位。与基于全局特征的指标（CLIP、DINO）或基于VLM的方法不同，VSM能够捕捉与姿态、尺度或上下文无关的细粒度外观变化。该包包含预训练的MTG模型权重、推理代码，以及展示特征提取、对应计算和不一致性评分的交互式playground notebook。训练代码、训练数据集生成管道和完整基准评估代码尚未发布。
+
+'
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: partial
+support_level: advanced
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- spearman_correlation
+- pearson_correlation
+- vsm_scores
+- mtg_inference_scores
+launch_profiles:
+- id: playground
+  label: Playground
+  description: '运行交互式playground notebook，加载预训练的MTG模型，对示例图像对执行推理，可视化解耦后的语义和视觉特征，并计算VSM评分以检测不一致区域。
+
+'
+- id: inference_eval
+  label: Inference Eval
+  description: '运行打包的评估脚本（eval_run_iter10_tta.py、eval_run_iter12_weighted_tta.py），计算VSM指标与测试数据集上Oracle评判之间的Spearman和Pearson相关性。注意，此快照中尚未发布完整训练管道和基准评估代码。
+
+'
+dataset_download:
+  primary_method: git_huggingface
+  sources:
+  - type: github
+    url: https://github.com/abdo-eldesokey/mind-the-glitch
+    note: 需要递归克隆；需配置Grounded-Segment-Anything子模块
+  - type: huggingface
+    url: https://huggingface.co/mind-the-glitch
+    note: 预训练模型权重和safetensors文件
+  notes:
+  - 使用递归克隆：git clone --recursive
+  - 如有需要，克隆后初始化子模块
+  - 设置HF_HOME环境变量用于模型缓存
+  - 部分模型下载可能需要HF token
+credential_requirements:
+  mode: optional_token
+  items:
+  - huggingface_token
+  notes:
+  - 部分受控模型访问可能需要HF_TOKEN；评估脚本使用匿名token
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    cuda_capability: '8.0'
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    cuda_capability: '8.6'
+environment:
+  python: '3.11'
+  cuda: '12.4'
+  pytorch: 2.5.1
+  flash_attn: null
+  key_packages:
+  - torch==2.5.1
+  - torchvision==0.20.1
+  - omegaconf
+  - hydra-core
+  - safetensors
+  - scipy
+  - Pillow
+  - numpy
+  - scikit-image
+  - GroundingDINO
+  - segment-anything
+  notes:
+  - 克隆后需单独安装Grounded-Segment-Anything
+  - 完整依赖列表见打包的requirements.txt
+  - 评估脚本已配置HF镜像以加快下载速度
+risk_flags:
+- inference_only
+- no_training_release
+- no_benchmark_eval_release
+risk_notes:
+- 当前快照仅提供推理功能
+- 训练代码和训练数据集生成管道尚未发布
+- 尚未包含完整基准评估框架
+- 静态代码审计确认可执行锚点存在
+- 本次打包过程中未执行运行时测试
+recommended_when: '当需要评估主体驱动图像生成中的主体一致性、比较VSM与现有指标（CLIP、DINO、LPIPS）、或利用视觉对应来定位生成主体内部的不一致区域时，可使用此基准。
+
+'
+not_recommended_when: '如果需要完整的训练管道、需要训练或微调MTG模型、或需要包含标准化测试集和Oracle对比的完整基准评估框架，则不应使用此基准。当前版本仅面向推理。
+
+'
+paper:
+  title: 'Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in Subject-Driven Generation'
+  authors:
+  - Abdelrahman Eldesokey
+  - Aleksandar Cvejic
+  - Bernard Ghanem
+  - Peter Wonka
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2509.21989
+  project_url: https://abdo-eldesokey.github.io/mind-the-glitch/
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.058_mindglitch.zip
+  archive_type: zip
+  local_dir_name: paper-58-MindGlitch
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.058_mindglitch.zip
+  sha256: 26309a90f48ab6a8b73a083a10ef1e7c2d3aed96f48a9bbf1bcf1cc2cd807b89
+  size_bytes: 3505243
+display:
+  palette_seed: mauve-graphite-correspond
+  art_style: visual-consistency
+  accent_priority: high
+image_path: ../image/058_aisb.t3.058_mindglitch.jpg
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+capability_tags:
+- computer_vision
+- diffusion_models
+- visual_correspondence
+- consistency_detection
+- subject_driven_generation
+- feature_disentanglement
+- vsm_metric
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.059_selfsupervised.yaml

@@ -0,0 +1,154 @@
+id: aisb.t3.059_selfsupervised
+name: SCAM Time-Series Self-Supervised Relabeling
+version: 0.1.0
+one_line: Self-supervised pseudo-label generation via reconstruction networks with
+  adaptive masking for time-series forecasting backbones.
+task_description: 'This benchmark implements SCAM (Self-Correction with Adaptive Mask),
+  a self-supervised relabeling framework for time-series forecasting. The method trains
+  a reconstruction network g(·; φ) on raw labels and generates pseudo-labels by approximating
+  intermediate network parameters during optimization. An adaptive mask identifies
+  overfitted raw components using the sharpness metric λ_max and selectively replaces
+  them with reconstruction-based pseudo-labels. Spectral Norm Regularization (SNR)
+  further suppresses overfitting. The benchmark supports multiple forecasting backbones
+  (MLP, CycleNet, iTransformer, PatchTST) and evaluates on 11 real-world datasets.
+  Users specify backbone via task config (e.g., task=scam_mlp, task=scam_patchtst)
+  in Hydra-based training pipeline.
+
+  '
+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:
+- avg_mse
+- avg_mae
+- selfsupervised_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run a single self-supervised forecasting evaluation route using a
+    packaged config. Executes src/eval.py with predefined hyperparameters for one
+    backbone model.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval
+  description: 'Run the complete SCAM self-supervised labeling and forecasting workflow.
+    Executes src/train.py with specified task config (scam_mlp, scam_cyclenet, scam_itransformer,
+    or scam_patchtst). Trains reconstruction network, applies adaptive masking, and
+    evaluates with SNR across all datasets.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: bundled
+    description: Datasets referenced in configs/data/ directory, downloaded via scripts
+  - type: url
+    description: External time-series datasets (ETT, weather, electricity, traffic)
+  notes:
+  - User must specify data paths in .env file (DATA_DIR variable) before execution
+  - See configs/data/ for dataset-specific configurations
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: GPU required for training reconstruction network and predictor models
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 24GB VRAM accommodates larger batch sizes for PatchTST/iTransformer experiments
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.6.0
+  flash_attn: null
+  key_packages:
+  - name: lightning
+    version: '>=2.0'
+    purpose: Training orchestration and distributed training support
+  - name: hydra-core
+    version: '>=1.3'
+    purpose: Configuration management and experiment routing
+  - name: omegaconf
+    version: '>=2.3'
+    purpose: YAML config parsing and overrides
+  - name: rootutils
+    version: '>=1.0'
+    purpose: Project root detection and PYTHONPATH setup
+  notes:
+  - Use environment.yml for conda environment creation: conda create --name <env>
+      --file environment.yml
+  - Python entry point uses lightning-hydra-template structure
+  - PROJECT_ROOT and DATA_DIR must be set in .env file before execution
+risk_flags:
+- requires_gpu
+- long_runtime
+risk_notes:
+- Training reconstruction network adds overhead compared to standard forecasting baselines
+- Full evaluation across all backbones (MLP, CycleNet, iTransformer, PatchTST) requires
+  multiple hours
+- 11 real-world datasets must be downloaded separately; disk space requirement scales
+  with dataset count
+recommended_when: 'Use this benchmark when evaluating self-supervised relabeling strategies,
+  studying pseudo-label quality effects on forecasting, or benchmarking SCAM against
+  standard supervised baselines. Particularly suited for research on label noise,
+  overfitting mitigation, and ensemble-style self-training in time series domains.
+
+  '
+not_recommended_when: 'Do not use if you need a fully deterministic single-model benchmark
+  without stochastic components. Not recommended when GPU resources are unavailable
+  or when baseline comparison without self-supervised components is required. Avoid
+  if reproducible single-run results are needed without the multi-stage SCAM pipeline.
+
+  '
+paper:
+  title: 'Not All Data are Good Labels: On the Self-supervised Labeling for Time Series
+    Forecasting'
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2502.14704
+  code_url: https://github.com/SuDIS-ZJU/SCAM
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.059_selfsupervised.zip
+  archive_type: zip
+  local_dir_name: paper-59-SelfSupervised
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.059_selfsupervised.zip
+  sha256: 4394ab146b3401b4b5dd318050d6b6feadca18a441a905e1ea4cbb49194d3de5
+  size_bytes: 167064
+display:
+  palette_seed: linen-blue-label
+  art_style: label-forecast
+  accent_priority: medium
+image_path: ../image/059_aisb.t3.059_selfsupervised.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- self_supervised_learning
+- forecasting
+- ensemble_methods
+- reconstruction_based_pseudolabels
+- adaptive_masking
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track

package/AISB/catalog/aisb.t3.059_selfsupervised.zh.yaml

@@ -0,0 +1,125 @@
+id: aisb.t3.059_selfsupervised
+name: SCAM 时序自监督重标注
+version: 0.1.0
+one_line: 通过带自适应掩码的重建网络生成自监督伪标签，用于时序预测骨干网络。
+task_description: |-
+  本基准测试实现了 SCAM（自适应掩码自纠正，Self-Correction with Adaptive Mask），这是一种用于时序预测的自监督重标注框架。该方法在原始标签上训练重建网络 g(·; φ)，并通过在优化过程中逼近网络中间参数来生成伪标签。自适应掩码使用尖锐度指标 λ_max 识别过拟合的原始分量，并选择性将其替换为基于重建的伪标签。谱范数正则化（SNR）进一步抑制过拟合。该基准测试支持多种预测骨干网络（MLP、CycleNet、iTransformer、PatchTST），并在 11 个真实数据集上进行评估。用户可通过任务配置指定骨干网络（如 task=scam_mlp、task=scam_patchtst），在基于 Hydra 的训练管道中运行。
+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:
+- avg_mse
+- avg_mae
+- selfsupervised_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: |-
+    使用打包配置运行单一自监督预测评估流程。通过预定义超参数执行 src/eval.py，对单个骨干模型进行评估。
+- id: full_train_eval
+  label: 完整训练 + 评估
+  description: |-
+    运行完整的 SCAM 自监督标注和预测工作流程。通过指定的任务配置（scam_mlp、scam_cyclenet、scam_itransformer 或 scam_patchtst）执行 src/train.py。训练重建网络，应用自适应掩码，并在所有数据集上使用 SNR 进行评估。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: bundled
+    description: configs/data/ 目录中引用的数据集，通过脚本下载
+  - type: url
+    description: 外部时序数据集（ETT、天气、电力、交通）
+  notes:
+  - 执行前，用户必须在 .env 文件中指定数据路径（DATA_DIR 变量）
+  - 详见 configs/data/ 中的数据集特定配置
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: 训练重建网络和预测器模型需要 GPU
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 24GB 显存可容纳更大批量大小，适用于 PatchTST/iTransformer 实验
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: 2.6.0
+  flash_attn: null
+  key_packages:
+  - name: lightning
+    version: '>=2.0'
+    purpose: 训练编排与分布式训练支持
+  - name: hydra-core
+    version: '>=1.3'
+    purpose: 配置管理与实验路由
+  - name: omegaconf
+    version: '>=2.3'
+    purpose: YAML 配置解析与覆盖
+  - name: rootutils
+    version: '>=1.0'
+    purpose: 项目根目录检测与 PYTHONPATH 配置
+  notes:
+  - 使用 environment.yml 创建 conda 环境：conda create --name <env> --file environment.yml
+  - Python 入口点使用 lightning-hydra-template 结构
+  - 执行前必须在 .env 文件中设置 PROJECT_ROOT 和 DATA_DIR
+risk_flags:
+- requires_gpu
+- long_runtime
+risk_notes:
+- 训练重建网络相比标准预测基线会产生额外开销
+- 在所有骨干网络（MLP、CycleNet、iTransformer、PatchTST）上进行完整评估需要数小时
+- 11 个真实数据集需要单独下载；磁盘空间需求随数据集数量增长
+recommended_when: |-
+  当评估自监督重标注策略、研究伪标签质量对预测的影响，或将 SCAM 与标准监督基线进行基准对比时使用本基准测试。特别适用于时序领域中标签噪声、过拟合缓解和集成式自训练的研究。
+not_recommended_when: |-
+  如需完全确定性的单一模型基准测试（不含随机成分），请勿使用。当 GPU 资源不可用或需要不带自监督成分的基线对比时，不建议使用。如需可复现的单次运行结果而无需多阶段 SCAM 流程，请避免使用。
+paper:
+  title: 'Not All Data are Good Labels: On the Self-supervised Labeling for Time Series
+    Forecasting'
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2502.14704
+  code_url: https://github.com/SuDIS-ZJU/SCAM
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.059_selfsupervised.zip
+  archive_type: zip
+  local_dir_name: paper-59-SelfSupervised
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.059_selfsupervised.zip
+  sha256: 4394ab146b3401b4b5dd318050d6b6feadca18a441a905e1ea4cbb49194d3de5
+  size_bytes: 167064
+display:
+  palette_seed: linen-blue-label
+  art_style: 标签预测
+  accent_priority: medium
+image_path: ../image/059_aisb.t3.059_selfsupervised.jpg
+capability_tags:
+- research_code_optimization
+- time_series_forecasting
+- self_supervised_learning
+- forecasting
+- ensemble_methods
+- reconstruction_based_pseudolabels
+- adaptive_masking
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track