@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.007_reasoningiqa.yaml

@@ -0,0 +1,169 @@
+schema_version: 1
+id: aisb.t3.007_reasoningiqa
+name: 'RALI: Reasoning-Aligned Lightweight Image Quality Assessment'
+version: 0.1.0
+one_line: Evaluate a CLIP-based lightweight IQA scorer (RALI) distilled from reasoning-heavy
+  RL-trained multimodal models on standard image quality benchmarks using PLCC, SRCC,
+  and MAE.
+task_description: 'This benchmark evaluates RALI, a lightweight no-reference image
+  quality assessment (IQA) model that uses CLIP-based contrastive alignment to match
+  the generalization of much larger reasoning-based IQA models (e.g., Q-Insight) at
+  ~4% of their parameters and inference time. The execution route loads pretrained
+  RALI checkpoints (CLIP vision encoder, PCA transform, prototype scorer), runs inference
+  on IQA benchmark datasets (KONIQ, KADID10K, SPAQ, etc.), and computes Pearson Linear
+  Correlation Coefficient (PLCC), Spearman Rank Correlation Coefficient (SRCC), and
+  Mean Absolute Error (MAE) against ground-truth subjective quality scores. The primary
+  entry point is eval_json.sh / eval_json.py, which scores images from JSON-defined
+  test splits and writes per-image predictions to a JSON log. A demo script (demo_rali_score.py)
+  provides single-image scoring. No training is required; all checkpoints are pretrained.
+  The benchmark tests whether an agent can correctly set up the environment, download
+  external checkpoints and dataset images, configure paths, and reproduce the paper''s
+  reported correlation metrics.
+
+  '
+capability_tags:
+- research_code_optimization
+- computer_vision
+- image_quality_assessment
+- representation_learning
+- evaluation
+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
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- plcc
+- srcc
+- mae
+launch_profiles:
+- id: quick_check
+  label: Quick Check
+  description: 'Run demo_rali_score.py on a single bundled image to verify that the
+    checkpoint loads, PCA and prototype scorer initialize, and a quality score is
+    produced. No dataset download required.
+
+    '
+- id: benchmark_eval
+  label: Benchmark Eval
+  description: 'Run eval_json.sh on one or more IQA benchmark test splits (e.g., KADID10K,
+    KONIQ, SPAQ). Requires downloading pretrained RALI checkpoints from HuggingFace
+    and the Data-DeQA-Score dataset images. Produces per-image prediction JSONs and
+    computes PLCC, SRCC, MAE.
+
+    '
+dataset_download:
+  primary_method: huggingface
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/ByteDance/Q-Insight/tree/main/RALI
+    access: public
+    note: 'RALI pretrained checkpoints (ckpt.pt, pca.pkl, basis.npz, and CLIP fine-tuned
+      model under best/). Place under ./checkpoints/.
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/zhiyuanyou/Data-DeQA-Score/tree/main
+    access: public
+    note: 'IQA benchmark meta files and source images for KONIQ, KADID10K, SPAQ, and
+      others. Place under ./Data-DeQA-Score/. Image directories may total several
+      GB.
+
+      '
+  notes:
+  - Checkpoint download is ~1-2 GB. Dataset images vary by benchmark; full Data-DeQA-Score
+    may exceed 10 GB.
+  - The snapshot includes directory stubs (Data-DeQA-Score/KADID10K, Data-DeQA-Score/KONIQ)
+    but not the actual images.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All checkpoints and datasets are publicly available on HuggingFace without authentication.
+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: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - transformers
+  - torch
+  - joblib
+  - numpy
+  - Pillow
+  - tqdm
+  notes:
+  - setup.sh installs all dependencies. See bundled README and requirements for the
+    full dependency set.
+  - Uses CLIPModel and CLIPProcessor from HuggingFace transformers.
+  - Exact Python/CUDA/PyTorch versions are not pinned; any recent compatible set should
+    work.
+risk_flags:
+- external_checkpoint_download
+- external_dataset_download
+- no_training_in_snapshot
+risk_notes:
+- Pretrained RALI checkpoints must be downloaded from HuggingFace (ByteDance/Q-Insight/RALI)
+  before any evaluation can run.
+- Dataset images are not bundled; only JSON meta files and directory stubs are included.
+  Images must be fetched from Data-DeQA-Score on HuggingFace.
+- No training code is included in the snapshot; this is inference/evaluation only.
+- No benchmark execution was performed during the packaging pass; metric values have
+  not been verified at runtime.
+recommended_when: 'Use this benchmark when you want a vision evaluation task with
+  pretrained checkpoints and straightforward image-benchmark execution. Good for testing
+  an agent''s ability to download external resources, configure paths, and run a self-contained
+  PyTorch inference pipeline that produces standard IQA correlation metrics.
+
+  '
+not_recommended_when: 'Do not use this if you need a large-scale training task, a
+  video quality benchmark, or a task that exercises long-horizon multi-stage model
+  training. Also not suitable if you want a fully self-contained benchmark with no
+  external downloads.
+
+  '
+paper:
+  title: 'Reasoning as Representation: Rethinking Visual Reinforcement Learning in
+    Image Quality Assessment'
+  venue: ICLR 2026 Oral
+  year: 2026
+  url: https://arxiv.org/abs/2510.11369
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.007_reasoningiqa.zip
+  archive_type: zip
+  local_dir_name: paper-7-ReasoningIQA
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.007_reasoningiqa.zip
+  sha256: 91b252c7f1616ea5b8889a80d6c6ba001c92f6b2f5db2998c133f533cdcc6bdf
+  size_bytes: 857520
+commercial:
+  annual_fee: null
+display:
+  palette_seed: cobalt-pearl-contrast
+  art_style: crisp-editorial
+  accent_priority: medium
+image_path: ../image/007_aisb.t3.007_reasoningiqa.jpg

package/AISB/catalog/aisb.t3.007_reasoningiqa.zh.yaml

@@ -0,0 +1,133 @@
+schema_version: 1
+id: aisb.t3.007_reasoningiqa
+name: 'RALI: 基于对齐推理的轻量级图像质量评估'
+version: 0.1.0
+one_line: 评估基于CLIP的轻量级IQA评分模型（RALI），该模型从推理密集型RL训练的多模态模型中蒸馏而来，在标准图像质量基准测试中使用PLCC、SRCC和MAE进行评估。
+task_description: '本基准测试评估RALI，这是一款轻量级无参考图像质量评估（IQA）模型，采用基于CLIP的对比对齐技术，在仅使用约4%参数和推理时间的前提下，达到与大型推理驱动IQA模型（如Q-Insight）相当的泛化能力。执行路由加载预训练的RALI检查点（CLIP视觉编码器、PCA变换、原型评分器），在IQA基准数据集（KONIQ、KADID10K、SPAQ等）上进行推理，并计算皮尔逊线性相关系数（PLCC）、斯皮尔曼等级相关系数（SRCC）和平均绝对误差（MAE）与主观质量评分ground truth的对比结果。主要入口点是eval_json.sh / eval_json.py，对JSON定义的测试分割中的图像进行评分，并将每张图像的预测结果写入JSON日志。演示脚本（demo_rali_score.py）提供单图像评分功能。无需训练，所有检查点均为预训练。本基准测试检验智能体能否正确设置环境、下载外部检查点和数据集图像、配置路径，并复现论文中报告的相关性指标。
+
+  '
+capability_tags:
+- research_code_optimization
+- computer_vision
+- image_quality_assessment
+- representation_learning
+- evaluation
+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
+time_band: 2-6h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- plcc
+- srcc
+- mae
+launch_profiles:
+- id: quick_check
+  label: 快速检查
+  description: '运行demo_rali_score.py对单个捆绑图像进行评分，验证检查点加载、PCA和原型评分器初始化，并产生质量评分。无需下载数据集。
+
+    '
+- id: benchmark_eval
+  label: 基准评估
+  description: '在某一个或多个IQA基准测试分割（如KADID10K、KONIQ、SPAQ）上运行eval_json.sh。需要从HuggingFace下载预训练的RALI检查点和Data-DeQA-Score数据集图像。生成每张图像的预测JSON，并计算PLCC、SRCC、MAE。
+
+    '
+dataset_download:
+  primary_method: huggingface
+  sources:
+  - kind: huggingface
+    url: https://huggingface.co/ByteDance/Q-Insight/tree/main/RALI
+    access: public
+    note: 'RALI预训练检查点（ckpt.pt、pca.pkl、basis.npz以及best/下的CLIP微调模型）。放置于./checkpoints/目录下。
+
+      '
+  - kind: huggingface
+    url: https://huggingface.co/datasets/zhiyuanyou/Data-DeQA-Score/tree/main
+    access: public
+    note: 'IQA基准测试元文件和KONIQ、KADID10K、SPAQ等数据集的源图像。放置于./Data-DeQA-Score/目录下。图像目录总计可能达数GB。
+
+      '
+  notes:
+  - 检查点下载约1-2 GB。数据集图像大小因基准测试而异，完整的Data-DeQA-Score可能超过10 GB。
+  - 快照包含目录存根（Data-DeQA-Score/KADID10K、Data-DeQA-Score/KONIQ），但不包含实际图像。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 所有检查点和数据集均可通过HuggingFace公开访问，无需身份验证。
+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: null
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - transformers
+  - torch
+  - joblib
+  - numpy
+  - Pillow
+  - tqdm
+  notes:
+  - setup.sh安装所有依赖项。参见捆绑的README和requirements以了解完整的依赖集合。
+  - 使用HuggingFace transformers中的CLIPModel和CLIPProcessor。
+  - 未固定精确的Python/CUDA/PyTorch版本，任何近期兼容版本组合均应可用。
+risk_flags:
+- external_checkpoint_download
+- external_dataset_download
+- no_training_in_snapshot
+risk_notes:
+- 必须从HuggingFace（ByteDance/Q-Insight/RALI）下载预训练的RALI检查点后才能运行任何评估。
+- 数据集图像未捆绑，仅包含JSON元文件和目录存根。图像必须从HuggingFace上的Data-DeQA-Score获取。
+- 快照中不包含训练代码，仅用于推理/评估。
+- 打包过程中未执行基准测试，运行时指标值未经核实。
+recommended_when: '当您需要一个视觉评估任务且具备预训练检查点和直接的图像基准执行时使用此基准测试。适合测试智能体下载外部资源、配置路径和运行自包含PyTorch推理管道以产生标准IQA相关性指标的能力。
+
+  '
+not_recommended_when: '如果需要大规模训练任务、视频质量基准测试或需要长时间多阶段模型训练的任务，请勿使用。也不适合需要完全自包含且无外部下载的基准测试。
+
+  '
+paper:
+  title: 'Reasoning as Representation: Rethinking Visual Reinforcement Learning in
+    Image Quality Assessment'
+  venue: ICLR 2026 Oral
+  year: 2026
+  url: https://arxiv.org/abs/2510.11369
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.007_reasoningiqa.zip
+  archive_type: zip
+  local_dir_name: paper-7-ReasoningIQA
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.007_reasoningiqa.zip
+  sha256: 91b252c7f1616ea5b8889a80d6c6ba001c92f6b2f5db2998c133f533cdcc6bdf
+  size_bytes: 857520
+commercial:
+  annual_fee: null
+display:
+  palette_seed: cobalt-pearl-contrast
+  art_style: crisp-editorial
+  accent_priority: medium
+image_path: ../image/007_aisb.t3.007_reasoningiqa.jpg

package/AISB/catalog/aisb.t3.008_meanflows.yaml

@@ -0,0 +1,188 @@
+schema_version: 1
+id: aisb.t3.008_meanflows
+name: Mean Flows for One-step Generative Modeling
+version: 0.1.0
+one_line: 'Train and evaluate a MeanFlow one-step generative model on CIFAR-10, measuring
+  FID-50K via a bundled evaluation script that requires a restored source tree and
+  checkpoint.
+
+  '
+task_description: 'This benchmark covers one-step image generation using the MeanFlow
+  framework, which learns an average velocity field (as opposed to instantaneous velocity
+  in standard Flow Matching) to enable single-NFE sampling. The packaged snapshot
+  includes an evaluation script (eval_cifar10.py) and config but is missing the core
+  MeanFlow training source tree (expected at /py-meanflow/meanflow/) and the pretrained
+  checkpoint (/checkpoints/cifar10_meanflow.pth). Before any benchmark run, agents
+  must restore the upstream MeanFlow codebase (model definitions, train_arg_parser,
+  model_configs) and obtain or retrain the CIFAR-10 checkpoint. Once restored, the
+  evaluation route generates 50,000 samples with a single forward pass per image (1-NFE)
+  using an EMA network, computes FID against CIFAR-10 training images via torchmetrics
+  FrechetInceptionDistance, and reports the scalar FID score. The paper reports FID
+  3.43 on ImageNet 256×256; the CIFAR-10 route here is a smaller-scale reproduction
+  target. Training uses a JVP-based loss with stop-gradient on the target, implemented
+  in JAX in the original paper but the evaluation script is PyTorch. No pre-training,
+  distillation, or curriculum learning is required by the method itself. The optimization
+  goal is to minimize FID while maintaining one-step (1-NFE) generation.
+
+  '
+capability_tags:
+- research_code_optimization
+- generative_modeling
+- diffusion_models
+- computer_vision
+- image_generation
+- flow_matching
+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: partial
+support_level: recovery
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- fid
+- generated_samples
+- evaluation_log
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: 'Inspect the preserved eval_cifar10.py, config.yaml, and metric contract
+    without running any generation. Useful for understanding the evaluation route
+    and planning restoration of the missing source tree.
+
+    '
+- id: restore_first
+  label: Restore First
+  description: 'Clone the upstream MeanFlow repository to provide model definitions
+    and train_arg_parser, obtain or retrain the CIFAR-10 checkpoint, then run the
+    full FID-50K evaluation pipeline via eval_cifar10.py.
+
+    '
+- id: train_and_eval
+  label: Train and Evaluate
+  description: 'Restore the full training pipeline, train MeanFlow on CIFAR-10 from
+    scratch using the JVP-based average-velocity loss, then evaluate FID-50K. Expect
+    multi-hour training on 2× A100 GPUs.
+
+    '
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - kind: torchvision
+    url: https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
+    access: public
+    note: 'CIFAR-10 training set (50K images, 32×32). The eval script expects pre-downloaded
+      data at /tmp/cifar10_data/cifar10_data with download=False. Use torchvision.datasets.CIFAR10(download=True)
+      to fetch automatically.
+
+      '
+  notes:
+  - CIFAR-10 is approximately 170 MB compressed, under 1 GB extracted.
+  - The pretrained checkpoint (cifar10_meanflow.pth) is not bundled and must be obtained
+    or retrained.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No API keys or credentials are needed. CIFAR-10 is freely downloadable.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 60
+    gpu_count: 2
+    gpu_vram_gb: 40
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 2
+    gpu_vram_gb: 80
+environment:
+  python: '3.10'
+  cuda: '12.1'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torchvision
+  - torchmetrics
+  - numpy
+  notes:
+  - The original paper implementation uses JAX with jax.jvp for the JVP-based training
+    loss; the bundled eval script uses PyTorch with torch.func.jvp.
+  - The local reproduction config used two A100 GPUs for FID-50K evaluation.
+  - The upstream MeanFlow source tree (model definitions, train_arg_parser, model_configs)
+    must be restored before execution.
+  - Batch size defaults to 128 for evaluation; adjust based on available VRAM.
+risk_flags:
+- missing_source_tree
+- missing_checkpoint
+- hardcoded_paths
+- no_prior_execution
+risk_notes:
+- The eval script hardcodes paths to /py-meanflow/meanflow/ and /checkpoints/cifar10_meanflow.pth
+  which do not exist in the snapshot.
+- The core MeanFlow model code (models/model_configs.py, train_arg_parser.py) is not
+  present in the local snapshot and must be sourced from the upstream repository.
+- No benchmark execution was performed during packaging; all metric values are unverified.
+- The README is absent from the current snapshot; rely on AGENTS.md, OPTIMIZATION.md,
+  and the paper for guidance.
+- The training route requires JAX if following the paper's implementation, but the
+  eval route is PyTorch-based, creating a potential framework mismatch.
+recommended_when: 'Use this benchmark when you want a compact, single-dataset image
+  generation task with a clear scalar FID target, one-step (1-NFE) sampling as the
+  core constraint, and interest in optimizing flow-based generative models without
+  distillation or curriculum learning. Good for agents that can handle code restoration
+  and dependency resolution.
+
+  '
+not_recommended_when: 'Do not use this if you cannot provision at least two modern
+  CUDA GPUs (≥40 GB VRAM each) for sample generation and FID scoring, or if you need
+  a fully self-contained ready-to-run benchmark without any source tree restoration.
+  Also avoid if you require ImageNet-scale results, as this snapshot targets CIFAR-10
+  only.
+
+  '
+paper:
+  title: Mean Flows for One-step Generative Modeling
+  authors:
+  - Zhengyang Geng
+  - Mingyang Deng
+  - Xingjian Bai
+  - J. Zico Kolter
+  - Kaiming He
+  affiliations:
+  - CMU
+  - MIT
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2505.13447
+  key_results:
+  - metric: FID
+    value: 3.43
+    dataset: ImageNet 256×256
+    condition: 1-NFE, trained from scratch, no distillation
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.008_meanflows.zip
+  archive_type: zip
+  local_dir_name: paper-8-MeanFlows
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.008_meanflows.zip
+  sha256: 3583357776f7b2d323b8a8170a88cb6559c1c90319e5b2947dc84ad1baf822e3
+  size_bytes: 100179
+commercial:
+  annual_fee: null
+display:
+  palette_seed: magenta-slate-ode
+  art_style: modern-lab
+  accent_priority: high
+image_path: ../image/008_aisb.t3.008_meanflows.jpg

package/AISB/catalog/aisb.t3.008_meanflows.zh.yaml

@@ -0,0 +1,140 @@
+schema_version: 1
+id: aisb.t3.008_meanflows
+name: 用于一步生成建模的均值流
+version: 0.1.0
+one_line: 在 CIFAR-10 上训练并评估 MeanFlow 一步生成模型，通过捆绑的评估脚本测量 FID-50K，该脚本需要恢复源代码树和检查点。
+
+task_description: 本基准测试涵盖使用 MeanFlow 框架进行的一步图像生成，该框架学习平均速度场（而非标准流匹配中的瞬时速度），以实现单次 NFE 采样。打包的快照包含评估脚本（eval_cifar10.py）和配置文件，但缺少核心 MeanFlow 训练源代码树（位于 /py-meanflow/meanflow/）和预训练检查点（/checkpoints/cifar10_meanflow.pth）。在任何基准测试运行之前，智能体必须恢复上游 MeanFlow 代码库（模型定义、train_arg_parser、model_configs）并获取或重新训练 CIFAR-10 检查点。恢复后，评估路线使用 EMA 网络通过单次前向传递（1-NFE）生成 50,000 个样本，通过 torchmetrics FrechetInceptionDistance 计算相对于 CIFAR-10 训练图像的 FID，并报告标量 FID 分数。论文报告 ImageNet 256×256 上 FID 为 3.43；此处 CIFAR-10 路线是较小规模的复现目标。训练使用基于 JVP 的损失函数和对目标的停止梯度，原始论文使用 JAX 实现，但评估脚本使用 PyTorch。该方法本身不需要预训练、蒸馏或课程学习。优化目标是在保持一步（1-NFE）生成的同时最小化 FID。
+
+capability_tags:
+- research_code_optimization
+- generative_modeling
+- diffusion_models
+- computer_vision
+- image_generation
+- flow_matching
+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: partial
+support_level: recovery
+cost_band: medium
+time_band: 6-24h
+difficulty: hard
+data_access: public
+primary_outputs:
+- fid
+- generated_samples
+- evaluation_log
+launch_profiles:
+- id: analysis_only
+  label: 仅分析
+  description: 检查保存的 eval_cifar10.py、config.yaml 和指标合约，无需运行任何生成过程。用于理解评估路线并规划恢复缺失的源代码树。
+
+- id: restore_first
+  label: 先恢复
+  description: 克隆上游 MeanFlow 仓库以提供模型定义和 train_arg_parser，获取或重新训练 CIFAR-10 检查点，然后通过 eval_cifar10.py 运行完整的 FID-50K 评估流程。
+
+- id: train_and_eval
+  label: 训练并评估
+  description: 恢复完整的训练流程，使用基于 JVP 的平均速度损失从头开始在 CIFAR-10 上训练 MeanFlow，然后评估 FID-50K。预计在 2× A100 GPU 上需要数小时训练。
+
+dataset_download:
+  primary_method: torchvision
+  sources:
+  - kind: torchvision
+    url: https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
+    access: public
+    note: CIFAR-10 训练集（50K 张图像，32×32）。评估脚本期望在 /tmp/cifar10_data/cifar10_data 处有预下载的数据，download=False。使用 torchvision.datasets.CIFAR10(download=True) 自动获取。
+
+  notes:
+  - CIFAR-10 压缩后约 170 MB，解压后不到 1 GB。
+  - 预训练检查点（cifar10_meanflow.pth）未捆绑，必须获取或重新训练。
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - 无需 API 密钥或凭证。CIFAR-10 可免费下载。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 60
+    gpu_count: 2
+    gpu_vram_gb: 40
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 2
+    gpu_vram_gb: 80
+environment:
+  python: '3.10'
+  cuda: '12.1'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - torchvision
+  - torchmetrics
+  - numpy
+  notes:
+  - 原始论文实现使用 JAX 和 jax.jvp 进行基于 JVP 的训练损失；捆绑的评估脚本使用 PyTorch 和 torch.func.jvp。
+  - 本地复现配置使用两枚 A100 GPU 进行 FID-50K 评估。
+  - 上游 MeanFlow 源代码树（模型定义、train_arg_parser、model_configs）必须在执行前恢复。
+  - 默认批量大小为 128 用于评估；根据可用 VRAM 进行调整。
+risk_flags:
+- missing_source_tree
+- missing_checkpoint
+- hardcoded_paths
+- no_prior_execution
+risk_notes:
+- 评估脚本硬编码了 /py-meanflow/meanflow/ 和 /checkpoints/cifar10_meanflow.pth 的路径，这些路径在快照中不存在。
+- 核心 MeanFlow 模型代码（models/model_configs.py、train_arg_parser.py）未出现在本地快照中，必须从上游仓库获取。
+- 打包期间未执行任何基准测试；所有指标值均未验证。
+- 当前快照中没有 README；请参考 AGENTS.md、OPTIMIZATION.md 和论文。
+- 如果遵循论文的实现，训练路线需要 JAX，但评估路线基于 PyTorch，可能存在框架不匹配。
+recommended_when: 当您需要一个紧凑的、单数据集的图像生成任务，具有明确的标量 FID 目标、一步（1-NFE）采样作为核心约束，并对优化基于流的生成模型（无需蒸馏或课程学习）感兴趣时，可使用此基准测试。适合能够处理代码恢复和依赖项解析的智能体。
+
+not_recommended_when: 如果无法为样本生成和 FID 评分配置至少两枚现代 CUDA GPU（每枚 ≥40 GB VRAM），或者需要一个完全自包含的、随时可运行的基准测试而无需任何源代码树恢复，请勿使用。如果需要 ImageNet 规模的结果，也应避免，因为此快照仅针对 CIFAR-10。
+
+paper:
+  title: Mean Flows for One-step Generative Modeling
+  authors:
+  - Zhengyang Geng
+  - Mingyang Deng
+  - Xingjian Bai
+  - J. Zico Kolter
+  - Kaiming He
+  affiliations:
+  - CMU
+  - MIT
+  venue: arXiv preprint
+  year: 2025
+  url: https://arxiv.org/abs/2505.13447
+  key_results:
+  - metric: FID
+    value: 3.43
+    dataset: ImageNet 256×256
+    condition: 1-NFE, 从头训练, 无蒸馏
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.008_meanflows.zip
+  archive_type: zip
+  local_dir_name: paper-8-MeanFlows
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.008_meanflows.zip
+  sha256: 3583357776f7b2d323b8a8170a88cb6559c1c90319e5b2947dc84ad1baf822e3
+  size_bytes: 100179
+commercial:
+  annual_fee: null
+display:
+  palette_seed: magenta-slate-ode
+  art_style: modern-lab
+  accent_priority: high
+image_path: ../image/008_aisb.t3.008_meanflows.jpg