@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.097_wassersteintl.yaml

@@ -0,0 +1,212 @@
+id: aisb.t3.097_wassersteintl
+name: Wasserstein Transfer Learning
+version: 0.1.0
+one_line: Optimize Wasserstein-based transfer estimators for distributional learning
+  across source and target populations using Fréchet regression in the Wasserstein
+  space.
+task_description: 'This packaged benchmark covers Wasserstein Transfer Learning (WaTL),
+  a novel transfer learning framework for regression models whose outputs are probability
+  distributions residing in the Wasserstein space. The method leverages knowledge
+  from multiple source domains to improve predictions in a target domain by intrinsically
+  incorporating the Wasserstein metric. The three-stage algorithm consists of: (1)
+  a weighted auxiliary estimator using global Fréchet regression for known informative
+  source subsets, (2) bias correction via gradient descent with cross-validated regularization,
+  and (3) projection to Wasserstein space enforcing monotonicity via OSQP solver.
+  Evaluation targets RMSPR (Root Mean Squared Prediction Risk) on mortality and physical
+  activity datasets, with per-test-point timing as a secondary metric.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- rmspr
+- time_ms_per_test_point
+- transfer_learning_report
+launch_profiles:
+- id: simulation
+  label: Simulation Route
+  description: 'Run the packaged simulation experiment (Section 5 of paper) to validate
+    WaTL asymptotic properties. Uses global Fréchet regression with synthetic distributional
+    data.
+
+    '
+  command: Rscript Simulation/Simulation.R <M> <n_t> <seed> <setting> <tau>
+  arguments:
+  - name: M
+    description: Grid size for quantile functions
+    example: '100'
+  - name: n_t
+    description: Target sample size
+    example: 200-800
+  - name: seed
+    description: Random seed for reproducibility
+    example: '42'
+  - name: setting
+    description: Data generation setting (1 or 2)
+    example: '1'
+  - name: tau
+    description: Source sample multiplier
+    example: 100 or 200
+- id: real_data
+  label: Real Data Route
+  description: 'Execute the real-world application with mortality or physical activity
+    data. Uses local Fréchet regression with OSQP monotonicity projection. Target
+    domain analysis with multiple source populations.
+
+    '
+  command: Rscript RealData/RealData.R <seed> <race> <M> <rate> <gender>
+  arguments:
+  - name: seed
+    description: Random seed for reproducibility
+    example: '42'
+  - name: race
+    description: Race index for target population (1=Black, 2=White)
+    example: '1'
+  - name: M
+    description: Grid size for quantile functions
+    example: '100'
+  - name: rate
+    description: Source data sampling rate (0-1)
+    example: '1.0'
+  - name: gender
+    description: Gender for stratification (0=Female, 1=Male)
+    example: '0'
+- id: mortality_experiment
+  label: Mortality Experiment (RMSPR)
+  description: 'Primary end-to-end reproduction path for RMSPR metric using UN World
+    Population Prospects 2015 data. Target: Developed countries (45). Sources: Developing
+    countries (156). Computes per-test-point timing.
+
+    '
+  command: Rscript mortality_experiment2.R
+  notes: 'Uses wpp2015 R package for demographic data. Requires internet for package
+    installation. Reproduces Table 1 (Appendix A).
+
+    '
+dataset_download:
+  primary_method: r_package
+  sources:
+  - name: wpp2015
+    description: UN World Population Prospects 2015 mortality data
+    package: wpp2015
+  - name: NHANES
+    description: National Health and Nutrition Examination Survey physical activity
+      data
+    package: NHANES
+  notes:
+  - wpp2015 package loads UN demographic data on first use
+  - NHANES physical activity intensity data loaded via RealDataFunc.R
+  - Both datasets are publicly available from NIH
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No authentication required for R package data sources
+  - Internet connection needed for package installation and data download
+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: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - name: wpp2015
+    description: UN World Population Prospects demographic data
+  - name: osqp
+    description: Quadratic programming solver for monotonicity projection
+  - name: Matrix
+    description: Sparse and dense matrix operations
+  - name: pracma
+    description: Practical numerical mathematics
+  - name: parallel
+    description: Parallel computation support
+  - name: NHANES
+    description: NHANES data retrieval
+  notes:
+  - CPU-only execution is sufficient for all routes
+  - GPU not required; workload is primarily statistical computation
+  - R-centric workflow; no Python or PyTorch dependencies
+  - OSQP solver used in Real Data route for quantile function monotonicity constraint
+risk_flags:
+- no_gpu_required
+- cpu_bound
+- r_only
+risk_notes:
+- Benchmark has no GPU requirement; resource costs are minimal
+- Execution time scales with grid size M and sample sizes
+- OSQP solver convergence may vary with data dimensionality
+recommended_when: 'Use this benchmark when you need a statistics-heavy transfer-learning
+  task that operates on distributional data in the Wasserstein space. Suitable for
+  evaluating methods that predict probability distributions from covariates using
+  optimal transport geometry. Ideal when working with mortality curves, physical activity
+  distributions, or other distributional response data where source domain knowledge
+  can improve target predictions. R-based execution makes it accessible without GPU
+  infrastructure.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you need neural-network-based
+  benchmarks, deep learning workflows, or CUDA-heavy training. Not suitable for tasks
+  requiring GPU acceleration or PyTorch-based implementations. Avoid if your research
+  focuses on scalar/vector responses rather than distributional outputs in non-Euclidean
+  spaces.
+
+  '
+paper:
+  title: Wasserstein Transfer Learning
+  authors:
+  - Kaicheng Zhang (Zhejiang University)
+  - Sinian Zhang (University of Minnesota)
+  - Doudou Zhou (National University of Singapore)
+  - Yidong Zhou (UC Davis)
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2505.17404
+  github: https://github.com/h7nian/WaTL
+display:
+  palette_seed: teal-stone-wasserstein
+  art_style: statistical-transport
+  accent_priority: medium
+image_path: ../image/097_aisb.t3.097_wassersteintl.jpg
+capability_tags:
+- transfer_learning
+- wasserstein_geometry
+- distributional_learning
+- statistical_modeling
+- optimal_transport
+- frechet_regression
+- non_euclidean_output
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.097_wassersteintl.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.097_wassersteintl.zip
+  archive_type: zip
+  sha256: 7878e684e20e96ca4ea5aa2ea5f33a4772d22a14b5bf6ab275a68664de302887
+  size_bytes: 104086

package/AISB/catalog/aisb.t3.097_wassersteintl.zh.yaml

@@ -0,0 +1,169 @@
+id: aisb.t3.097_wassersteintl
+name: Wasserstein 迁移学习
+version: 0.1.0
+one_line: 使用 Wasserstein 空间中的 Fréchet 回归优化基于 Wasserstein 的迁移估计器，以实现源群体和目标群体之间的分布式学习。
+task_description: 该打包基准涵盖 Wasserstein 迁移学习 (WaTL)，这是一种用于回归模型的新型迁移学习框架，其输出是位于 Wasserstein 空间中的概率分布。该方法通过内在地结合 Wasserstein 度量，利用来自多个源域的知识来改进目标域中的预测。该三阶段算法包括：(1) 对已知信息源子集使用全局 Fréchet 回归的加权辅助估计器，(2) 通过交叉验证正则化的梯度下降进行偏差校正，以及 (3) 通过 OSQP 求解器投影到 Wasserstein 空间以强制单调性。评估目标是死亡率和身体活动数据集的 RMSPR（均方根预测风险），每个测试点的时间作为次要指标。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- rmspr
+- time_ms_per_test_point
+- transfer_learning_report
+launch_profiles:
+- id: simulation
+  label: 模拟路线
+  description: 运行打包的模拟实验（论文第 5 节）以验证 WaTL 渐近属性。将全局 Fréchet 回归与合成分布数据结合使用。
+  command: Rscript Simulation/Simulation.R <M> <n_t> <seed> <setting> <tau>
+  arguments:
+  - name: M
+    description: 分位数函数的网格大小
+    example: '100'
+  - name: n_t
+    description: 目标样本量
+    example: 200-800
+  - name: seed
+    description: 用于再现性的随机种子
+    example: '42'
+  - name: setting
+    description: 数据生成设置（1或2）
+    example: '1'
+  - name: tau
+    description: 源样本乘数
+    example: 100 or 200
+- id: real_data
+  label: 真实数据路由
+  description: 使用死亡率或体力活动数据执行现实世界的应用程序。使用带有 OSQP 单调性投影的局部 Fréchet 回归。具有多个源群体的目标域分析。
+  command: Rscript RealData/RealData.R <seed> <race> <M> <rate> <gender>
+  arguments:
+  - name: seed
+    description: 用于再现性的随机种子
+    example: '42'
+  - name: race
+    description: 目标人群的种族指数（1=黑人，2=白人）
+    example: '1'
+  - name: M
+    description: 分位数函数的网格大小
+    example: '100'
+  - name: rate
+    description: 源数据采样率（0-1）
+    example: '1.0'
+  - name: gender
+    description: 分层性别（0=女性，1=男性）
+    example: '0'
+- id: mortality_experiment
+  label: 死亡率实验 (RMSPR)
+  description: 使用联合国 2015 年世界人口展望数据的 RMSPR 指标的主要端到端复制路径。目标：发达国家（45）。资料来源：发展中国家（156）。计算每个测试点的时序。
+  command: Rscript mortality_experiment2.R
+  notes: 使用 wpp2015 R 包获取人口统计数据。需要互联网来安装软件包。重现表 1（附录 A）。
+dataset_download:
+  primary_method: r_package
+  sources:
+  - name: wpp2015
+    description: UN World Population Prospects 2015 mortality data
+    package: wpp2015
+  - name: NHANES
+    description: National Health and Nutrition Examination Survey physical activity data
+    package: NHANES
+  notes:
+  - wpp2015 软件包首次使用时加载联合国人口统计数据
+  - NHANES 身体活动强度数据通过 RealDataFunc.R 加载
+  - 这两个数据集均可从 NIH 公开获得
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - R包数据源无需身份验证
+  - 软件包安装和数据下载需要互联网连接
+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: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - name: wpp2015
+    description: UN World Population Prospects demographic data
+  - name: osqp
+    description: Quadratic programming solver for monotonicity projection
+  - name: Matrix
+    description: Sparse and dense matrix operations
+  - name: pracma
+    description: Practical numerical mathematics
+  - name: parallel
+    description: Parallel computation support
+  - name: NHANES
+    description: NHANES data retrieval
+  notes:
+  - 仅 CPU 执行足以满足所有路由的需要
+  - 不需要 GPU；工作量主要是统计计算
+  - 以 R 为中心的工作流程；没有 Python 或 PyTorch 依赖项
+  - OSQP 求解器在真实数据路径中用于分位数函数单调性约束
+risk_flags:
+- no_gpu_required
+- cpu_bound
+- r_only
+risk_notes:
+- Benchmark 没有 GPU 要求；资源成本极低
+- 执行时间与网格大小 M 和样本大小相关
+- OSQP 求解器收敛可能随数据维度而变化
+recommended_when: 当您需要对 Wasserstein 空间中的分布数据进行操作的统计量大的迁移学习任务时，请使用此基准。适用于评估使用最佳传输几何从协变量预测概率分布的方法。非常适合处理死亡率曲线、体力活动分布或其他分布响应数据，其中源领域知识可以改进目标预测。基于 R 的执行使其无需 GPU 基础设施即可访问。
+not_recommended_when: 如果您需要基于神经网络的基准测试、深度学习工作流程或 CUDA 密集型训练，请勿使用此基准测试。不适合需要 GPU 加速或基于 PyTorch 的实现的任务。如果您的研究重点是标量/向量响应而不是非欧几里得空间中的分布输出，请避免使用。
+paper:
+  title: Wasserstein Transfer Learning
+  authors:
+  - Kaicheng Zhang (Zhejiang University)
+  - Sinian Zhang (University of Minnesota)
+  - Doudou Zhou (National University of Singapore)
+  - Yidong Zhou (UC Davis)
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2505.17404
+  github: https://github.com/h7nian/WaTL
+display:
+  palette_seed: teal-stone-wasserstein
+  art_style: statistical-transport
+  accent_priority: medium
+image_path: ../image/097_aisb.t3.097_wassersteintl.jpg
+capability_tags:
+- transfer_learning
+- wasserstein_geometry
+- distributional_learning
+- statistical_modeling
+- optimal_transport
+- frechet_regression
+- non_euclidean_output
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.097_wassersteintl.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.097_wassersteintl.zip
+  archive_type: zip
+  sha256: 7878e684e20e96ca4ea5aa2ea5f33a4772d22a14b5bf6ab275a68664de302887
+  size_bytes: 104086

package/AISB/catalog/aisb.t3.098_xmahalanobis.yaml

@@ -0,0 +1,171 @@
+id: aisb.t3.098_xmahalanobis
+name: 'X-Mahalanobis: Transformer Feature Mixing for Reliable OOD Detection'
+version: 0.1.0
+one_line: Adaptive transformer layer fusion with Mahalanobis distance scoring for
+  state-of-the-art out-of-distribution detection.
+task_description: 'This packaged benchmark covers transformer-based out-of-distribution
+  (OOD) detection through an adaptive fusion module that dynamically assigns importance
+  weights to representations learned by each Transformer layer. The method detects
+  OOD samples using Mahalanobis distance scoring across multiple intermediate layers,
+  moving beyond reliance on final-layer features alone. The benchmark supports parameter-efficient
+  fine-tuning (PEFT) strategies including AdaptFormer, LoRA, Adapter, Visual Prompt
+  Tuning (VPT), bias tuning, and full fine-tuning. Evaluation is conducted on class-balanced
+  (CIFAR100) and long-tailed (ImageNet-LT) in-distribution classification tasks against
+  nine diverse OOD datasets, reporting AUROC and FPR95 metrics.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- auroc
+- fpr95
+- ood_detection_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run OOD detection evaluation on a prepared checkpoint using main_test.py.
+    Requires trained model weights and config alignment with training setup. Suitable
+    for evaluating baselines or comparing fine-tuned checkpoints.
+
+    '
+  estimated_time: 1-2h
+  compute_profile: quick_eval
+- id: train_and_test
+  label: Train + Test
+  description: 'Full PEFT training workflow on CIFAR100 or ImageNet-LT followed by
+    OOD evaluation. Supports AdaptFormer (default), LoRA, Adapter, VPT, bias tuning,
+    and full fine-tuning. Trains ViT-B/16 (IN21K pretrained or CLIP pretrained) and
+    evaluates against multiple OOD datasets.
+
+    '
+  estimated_time: 6-18h
+  compute_profile: full_training
+dataset_download:
+  primary_method: bundled
+  sources:
+  - cifar100: torchvision datasets (auto-download)
+  - imagenet_lt: requires manual download from ImageNet, LT split preconfigured in
+      configs
+  - ood_datasets: preconfigured in evaluation pipeline (e.g., SVHN, LSUN, iNaturalist)
+  notes:
+  - CIFAR100 downloads automatically via torchvision
+  - ImageNet-LT requires manual download from ImageNet official source
+  - OOD test datasets (TIN, SVHN, LSUN, etc.) handled by evaluation script
+  - Data configs located in configs/data/ (cifar100.yaml, imagenet_lt.yaml, imagenet.yaml)
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - ImageNet download may require academic registration
+  - All other datasets are publicly accessible or auto-downloaded
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: Supports single-GPU execution for quick_eval and small-scale training
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: '24GB VRAM recommended for batch_size 64 with full PEFT methods on ImageNet-LT.
+      AdaptFormer training at lr 0.1 for 20 epochs on ImageNet-LT with CLIP ViT-B/16
+      is the most demanding configuration.
+
+      '
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages:
+  - torch>=2.0
+  - torchvision
+  - numpy
+  - scikit-learn
+  - scipy
+  - timm
+  - open_clip_torch
+  notes:
+  - See bundled requirements.txt or README for full dependency set
+  - open_clip_torch required for CLIP ViT-B/16 model support
+  - timm required for IN21K ViT-B/16 pretrained models
+risk_flags:
+- long_runtime
+- gpu_memory_sensitive
+- config_sensitivity
+risk_notes:
+- Full training on ImageNet-LT with 20 epochs may take 12-18 hours
+- Batch size 64 on 24GB GPU requires AdaptFormer PEFT; full fine-tuning may OOM
+- Config alignment between train and test is critical; mismatched configs produce
+  invalid results
+- PEFT method flags (adaptformer, lora, adapter, etc.) must be explicitly set per
+  experiment
+- Seed setting via --seed flag available but not enforced; reproducibility requires
+  manual seed setting
+recommended_when: 'Use this benchmark when developing transformer-based reliability
+  and OOD detection methods. Particularly suited for research on: (1) adaptive layer
+  fusion strategies, (2) Mahalanobis-based OOD scoring, (3) parameter-efficient fine-tuning
+  for out-of-distribution generalization, (4) comparing shallow vs. deep transformer
+  representations, and (5) benchmarking OOD detection under class-imbalanced (long-tailed)
+  distributions.
+
+  '
+not_recommended_when: 'Do not use this benchmark for: (1) text-only reliability tasks
+  (OOD detection here is vision-only), (2) non-transformer architectures (pure CNN-based
+  methods), (3) tasks requiring OOD detection without fine-tuning (zero-shot OOD evaluation),
+  or (4) resource-constrained environments without GPU access.
+
+  '
+paper:
+  title: 'X-Mahalanobis: Transformer Feature Mixing for Reliable OOD Detection'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.09128
+  notes: 'NeurIPS 2025 publication; benchmark packages official implementation from
+    the paper. Experiments demonstrate SOTA OOD detection across nine OOD datasets.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.098_xmahalanobis.zip
+  archive_type: zip
+  local_dir_name: paper-98-XMahalanobis
+  notes: 'Bundled archive includes full training/test code, configs, utility modules,
+    and pretrained model loading support for CLIP and IN21K ViT variants.
+
+    '
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.098_xmahalanobis.zip
+  sha256: c65b3f81a78754d84bff9a76b8a0c13e180f6b3b6d379801647ed189b1f4741e
+  size_bytes: 1420263
+display:
+  palette_seed: cobalt-silver-mahalanobis
+  art_style: reliability-vision
+  accent_priority: high
+image_path: ../image/098_aisb.t3.098_xmahalanobis.jpg
+capability_tags:
+- ood_detection
+- transformers
+- computer_vision
+- reliable_ml
+- parameter_efficient_finetuning
+- layer_fusion
+- mahalanobis_distance
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track

package/AISB/catalog/aisb.t3.098_xmahalanobis.zh.yaml

@@ -0,0 +1,127 @@
+id: aisb.t3.098_xmahalanobis
+name: X-Mahalanobis：用于可靠 OOD 检测的变压器特征混合
+version: 0.1.0
+one_line: 自适应变换层融合与马哈拉诺比斯距离评分，用于最先进的分布外检测。
+task_description: 该打包基准测试通过自适应融合模块涵盖基于 Transformer 的分布外 (OOD) 检测，该模块动态地将重要性权重分配给每个 Transformer 层学习的表示。该方法使用跨多个中间层的马哈拉诺比斯距离评分来检测 OOD 样本，不再仅仅依赖最终层特征。该基准测试支持参数高效微调 (PEFT) 策略，包括 AdaptFormer、LoRA、Adapter、视觉提示调整 (VPT)、偏差调整和完全微调。针对九个不同的 OOD 数据集对类平衡 (CIFAR100) 和长尾 (ImageNet-LT) 分布内分类任务进行评估，报告 AUROC 和 FPR95 指标。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- auroc
+- fpr95
+- ood_detection_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 使用 main_test.py 在准备好的检查点上运行 OOD 检测评估。需要经过训练的模型权重和配置与训练设置保持一致。适合评估基线或比较微调的检查点。
+  estimated_time: 1-2h
+  compute_profile: quick_eval
+- id: train_and_test
+  label: 训练+测试
+  description: CIFAR100 或 ImageNet-LT 上的完整 PEFT 训练工作流程，然后进行 OOD 评估。支持AdaptFormer（默认）、LoRA、Adapter、VPT、偏置调整和全面微调。训练 ViT-B/16（IN21K 预训练或 CLIP 预训练）并针对多个 OOD 数据集进行评估。
+  estimated_time: 6-18h
+  compute_profile: full_training
+dataset_download:
+  primary_method: bundled
+  sources:
+  - cifar100: torchvision datasets (auto-download)
+  - imagenet_lt: requires manual download from ImageNet, LT split preconfigured in configs
+  - ood_datasets: preconfigured in evaluation pipeline (e.g., SVHN, LSUN, iNaturalist)
+  notes:
+  - CIFAR100 通过 torchvision 自动下载
+  - ImageNet-LT需要从ImageNet官方源手动下载
+  - 由评估脚本处理的 OOD 测试数据集（TIN、SVHN、LSUN 等）
+  - 数据配置位于 configs/data/ (cifar100.yaml, imagenet_lt.yaml, imagenet.yaml)
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - ImageNet 下载可能需要学术注册
+  - 所有其他数据集均可公开访问或自动下载
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: 支持单 GPU 执行 quick_eval 和小规模训练
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: 建议 batch_size 64 使用 24GB VRAM，在 ImageNet-LT 上使用完整的 PEFT 方法。 AdaptFormer 使用 CLIP ViT-B/16 在 ImageNet-LT 上以 lr 0.1 进行 20 个周期的训练是要求最高的配置。
+environment:
+  python: '3.9'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages:
+  - torch>=2.0
+  - torchvision
+  - numpy
+  - scikit-learn
+  - scipy
+  - timm
+  - open_clip_torch
+  notes:
+  - 请参阅捆绑的 requirements.txt 或 README 以获取完整的依赖项集
+  - CLIP ViT-B/16 模型支持需要 open_clip_torch
+  - IN21K ViT-B/16 预训练模型所需的 timm
+risk_flags:
+- long_runtime
+- gpu_memory_sensitive
+- config_sensitivity
+risk_notes:
+- 对 ImageNet-LT 进行 20 个 epoch 的完整训练可能需要 12-18 小时
+- 24GB GPU 上的批量大小 64 需要 AdaptFormer PEFT；完全微调可能会OOM
+- 训练和测试之间的配置对齐至关重要；不匹配的配置会产生无效结果
+- 每个实验必须明确设置 PEFT 方法标志（adaptformer、lora、adapter 等）
+- 通过 --seed 标志设置种子可用，但不强制执行；再现性需要手动播种
+recommended_when: 开发基于变压器的可靠性和 OOD 检测方法时使用此基准。特别适合研究：(1) 自适应层融合策略，(2) 基于 Mahalanobis 的 OOD 评分，(3) 针对分布外泛化的参数高效微调，(4) 比较浅层和深层变压器表示，以及 (5) 类不平衡（长尾）分布下的基准 OOD 检测。
+not_recommended_when: 请勿将此基准用于：(1) 纯文本可靠性任务（此处的 OOD 检测仅针对视觉）、(2) 非 Transformer 架构（纯基于 CNN 的方法）、(3) 需要无需微调的 OOD 检测的任务（零样本 OOD 评估）或 (4) 无法访问 GPU 的资源受限环境。
+paper:
+  title: 'X-Mahalanobis: Transformer Feature Mixing for Reliable OOD Detection'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2505.09128
+  notes: NeurIPS 2025 出版物；基准包官方实现来自论文。实验证明了跨九个 OOD 数据集的 SOTA OOD 检测。
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.098_xmahalanobis.zip
+  archive_type: zip
+  local_dir_name: paper-98-XMahalanobis
+  notes: 'Bundled archive includes full training/test code, configs, utility modules, and pretrained model loading support for CLIP and IN21K ViT variants.
+
+    '
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.098_xmahalanobis.zip
+  sha256: c65b3f81a78754d84bff9a76b8a0c13e180f6b3b6d379801647ed189b1f4741e
+  size_bytes: 1420263
+display:
+  palette_seed: cobalt-silver-mahalanobis
+  art_style: reliability-vision
+  accent_priority: high
+image_path: ../image/098_aisb.t3.098_xmahalanobis.jpg
+capability_tags:
+- ood_detection
+- transformers
+- computer_vision
+- reliable_ml
+- parameter_efficient_finetuning
+- layer_fusion
+- mahalanobis_distance
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track