@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.070_certifiedunlearning.yaml

@@ -0,0 +1,224 @@
+schema_version: 1
+id: aisb.t3.070_certifiedunlearning
+name: Certified Unlearning for Neural Networks
+version: 0.1.0
+one_line: Certified machine unlearning via privacy amplification by noisy fine-tuning,
+  with provable (ε,δ)-guarantees and no loss-function assumptions.
+task_description: 'This packaged benchmark implements and evaluates certified machine
+  unlearning for neural networks (ICML 2025). The core task is to remove the influence
+  of a specified forget set from a trained model while preserving utility on the retain
+  set, subject to formal (ε,δ)-unlearning guarantees. The method—Privacy Amplification
+  by Iteration (PABI)—treats each noisy fine-tuning step on the retain data as a stochastic
+  post-processing operation, progressively amplifying privacy without accessing the
+  forget data. It imposes no smoothness or convexity assumptions on the loss function,
+  making it applicable to arbitrary deep networks. Four methods are benchmarked: PABI
+  (gradient clipping), DP-SGD with group privacy, Contractive Coefficients (model
+  clipping), and DP-Baseline (output perturbation), alongside full Retrain. Experiments
+  cover CIFAR-10, CIFAR-100, and MNIST across four figure-level ablations: privacy
+  budget curves, convergence curves, DP-SGD comparisons, and ε-sweep analysis. The
+  primary efficiency metric is epochs_to_50pct_val_acc measured during the unlearning
+  phase. The framework uses JAX/Flax for all training and unlearning logic; feature
+  extraction uses a ResNet-18 backbone. Multi-GPU execution is supported via run_exp.py.
+  No benchmark execution was performed during this packaging pass—runtime validation
+  is required before trusting metric values.
+
+  '
+capability_tags:
+- machine_unlearning
+- certified_privacy
+- differential_privacy
+- privacy_amplification
+- neural_networks
+- research_code_optimization
+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: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- epochs_to_50pct_val_acc
+- epochs_to_45pct_val_acc
+- epochs_to_40pct_val_acc
+- epochs_to_35pct_val_acc
+- epochs_to_30pct_val_acc
+- unlearning_checkpoint
+- certification_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged certified unlearning evaluation route using a single
+    pre-configured experiment YAML (e.g., from dp_sgd_runs/). Produces a single certification
+    report and efficiency metrics. Fastest path to verifying the environment and metric
+    surface.
+
+    '
+- id: full_unlearning
+  label: Full Unlearning
+  description: 'Run the complete privacy-aware unlearning training and certification
+    workflow across all four experiment suites (budget_curve_runs, convergence_curve_runs,
+    dp_sgd_runs, eps_sweep_runs). Reproduces all paper figures. Requires multi-GPU
+    execution via run_exp.py.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: CIFAR-10
+    notes: Auto-downloaded by JAX/datasets pipeline at runtime.
+  - name: CIFAR-100
+    notes: Auto-downloaded by JAX/datasets pipeline at runtime.
+  - name: MNIST
+    notes: Auto-downloaded by JAX/datasets pipeline at runtime.
+  - name: ResNet-18 feature cache
+    notes: 'Generated locally by running feature_extractor.py before experiments that
+      use cifar10_feature or cifar100_feature dataset configs.
+
+      '
+  notes:
+  - All datasets are standard public benchmarks; no manual download required unless
+    offline execution is needed.
+  - feature_extractor.py must be run once before transfer-learning experiment configs
+    are used.
+credential_requirements:
+  mode: optional
+  items:
+  - name: Weights & Biases (wandb)
+    notes: 'experiment.py calls wandb.init() at startup. Set WANDB_API_KEY or run
+      `wandb login` before execution. Offline mode can be enabled by setting WANDB_MODE=offline.
+
+      '
+  notes:
+  - No paid data credentials required. wandb is the only external service.
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: Sufficient for single-experiment quick_eval runs on MNIST or CIFAR-10.
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 2
+    gpu_vram_gb: 24
+    notes: 'Two GPUs allow run_exp.py to parallelize across experiment suites. Required
+      for full reproduction of all four figure-level ablations within the 6–24 h time
+      band.
+
+      '
+environment:
+  python: '3.8'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  framework: jax
+  key_packages:
+  - jax==0.5.2
+  - jaxlib==0.5.1
+  - jax-cuda12-pjrt==0.5.1
+  - jax-cuda12-plugin==0.5.1
+  - flax==0.10.4
+  - optax
+  - dp-accounting==0.4.4
+  - transformers==4.49.0
+  - datasets==2.21.0
+  - ml_collections==1.0.0
+  - wandb
+  - chex==0.1.89
+  - torch
+  - matplotlib==3.10.1
+  - jaxpruner @ git+https://github.com/google-research/jaxpruner@e7af4e4e4748782d6bc433249e359f74ffb845ba
+  notes:
+  - This benchmark uses JAX/Flax as its primary framework, not PyTorch. PyTorch is
+    present in requirements.txt only for DataLoader utilities.
+  - CUDA version is determined by jax-cuda12-* wheel; install JAX CUDA packages before
+    other dependencies.
+  - See requirements.txt for the full pinned dependency set.
+  - Python 3.8 is required; the JAX version used (0.5.2) does not support 3.12+.
+risk_flags:
+- jax_cuda_wheel_version_sensitivity
+- wandb_required_at_runtime
+- feature_extractor_prereq
+- no_runtime_execution_in_packaging_pass
+risk_notes:
+- 'JAX CUDA wheels are highly version-sensitive. Install jax-cuda12-pjrt and jax-cuda12-plugin
+  before jaxlib to avoid ABI mismatches.
+
+  '
+- 'experiment.py calls wandb.init() unconditionally; set WANDB_MODE=offline to prevent
+  a hard failure when no API key is present.
+
+  '
+- 'Experiments using cifar10_feature or cifar100_feature dataset configs will fail
+  if feature_extractor.py has not been run first.
+
+  '
+- 'No benchmark execution was performed during this packaging pass. All metric values
+  in metric_contract.json are provisional until runtime validation confirms them.
+  Do not trust metric surface from README alone.
+
+  '
+recommended_when: 'Use this benchmark when evaluating certified machine unlearning
+  methods on image-classification tasks (CIFAR-10, CIFAR-100, MNIST) with formal (ε,δ)
+  privacy guarantees. Ideal for comparing unlearning efficiency (epochs to reach a
+  validation accuracy threshold) across PABI, DP-SGD, contractive coefficient, and
+  output-perturbation baselines. Best suited for researchers who need assumption-free
+  certified unlearning applicable to non-convex deep networks.
+
+  '
+not_recommended_when: 'Do not use if you only need standard image-classification training
+  without privacy or unlearning constraints. Not suitable if your pipeline requires
+  PyTorch-native training loops—the core implementation is JAX/Flax only. Avoid if
+  GPU VRAM is below 12 GB or if a JAX CUDA environment cannot be established. Not
+  appropriate for large-scale language-model unlearning or concept unlearning tasks,
+  which require a different method family.
+
+  '
+paper:
+  title: Certified Unlearning for Neural Networks
+  authors:
+  - Anastasia Koloskova
+  - Youssef Allouah
+  - Animesh Jha
+  - Rachid Guerraoui
+  - Sanmi Koyejo
+  affiliations:
+  - Stanford University
+  - EPFL
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2506.06985
+  bibtex_key: koloskova2025certified
+  notes: 'Equal contribution from Koloskova and Allouah. Proceedings of the 42nd International
+    Conference on Machine Learning, Vancouver, Canada. PMLR 267.
+
+    '
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.070_certifiedunlearning.zip
+  archive_type: zip
+  local_dir_name: paper-70-CertifiedUnlearning
+  upstream_repo: https://github.com/stair-lab/certified-unlearning-neural-networks-icml-2025
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.070_certifiedunlearning.zip
+  sha256: 124e11876363b6ccc532c5d235697f96975e0e2883506083a5f9391b63cf1ab7
+  size_bytes: 13612960
+commercial:
+  annual_fee: null
+display:
+  palette_seed: silver-plum-unlearn
+  art_style: privacy-lab
+  accent_priority: high
+image_path: ../image/070_aisb.t3.070_certifiedunlearning.jpg

package/AISB/catalog/aisb.t3.070_certifiedunlearning.zh.yaml

@@ -0,0 +1,171 @@
+schema_version: 1
+id: aisb.t3.070_certifiedunlearning
+name: 神经网络认证遗忘
+version: 0.1.0
+one_line: 通过噪声微调实现隐私放大进行认证机器遗忘，提供可证明的(ε,δ)保证且无需损失函数假设。
+task_description: >
+  本基准测试包实现并评估了神经网络认证机器遗忘方法（ICML 2025）。核心任务是从已训练模型中移除指定遗忘集的影响，同时在保留集上保持效用，并满足正式的(ε,δ)遗忘保证。该方法——迭代隐私放大（PABI）——将保留数据上每个带噪声的微调步骤视为随机后处理操作，逐步放大隐私性而不访问遗忘数据。该方法对损失函数不做平滑性或凸性假设，适用于任意深度网络。基准测试涵盖四种方法：PABI（梯度裁剪）、基于组隐私的DP-SGD、收缩系数（模型裁剪）和DP基线（输出扰动），以及完整重训练作为对照。实验覆盖CIFAR-10、CIFAR-100和MNIST数据集，包含四组消融实验：隐私预算曲线、收敛曲线、DP-SGD对比和ε扫描分析。主要效率指标为遗忘阶段中epochs_to_50pct_val_acc。框架使用JAX/Flax实现所有训练和遗忘逻辑；特征提取使用ResNet-18骨干网络。支持通过run_exp.py进行多GPU执行。
+
+  本打包过程未执行任何基准测试运行——在使用指标数值前需进行运行时验证。
+
+capability_tags:
+- machine_unlearning
+- certified_privacy
+- differential_privacy
+- privacy_amplification
+- neural_networks
+- research_code_optimization
+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: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- epochs_to_50pct_val_acc
+- epochs_to_45pct_val_acc
+- epochs_to_40pct_val_acc
+- epochs_to_35pct_val_acc
+- epochs_to_30pct_val_acc
+- unlearning_checkpoint
+- certification_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: >
+    使用单个预配置实验YAML（例如来自dp_sgd_runs/）运行一条封装的认证遗忘评估路线。生成一份认证报告和效率指标。是验证环境和指标表面的最快路径。
+
+- id: full_unlearning
+  label: 完整遗忘
+  description: >
+    跨所有四组实验套件（budget_curve_runs、convergence_curve_runs、dp_sgd_runs、eps_sweep_runs）运行完整的隐私感知遗忘训练和认证工作流。复现论文所有图表。需要通过run_exp.py进行多GPU执行。
+
+dataset_download:
+  primary_method: mixed
+  sources:
+  - name: CIFAR-10
+    notes: 由JAX/datasets管道在运行时自动下载。
+  - name: CIFAR-100
+    notes: 由JAX/datasets管道在运行时自动下载。
+  - name: MNIST
+    notes: 由JAX/datasets管道在运行时自动下载。
+  - name: ResNet-18特征缓存
+    notes: 通过在实验前运行feature_extractor.py在本地生成，用于cifar10_feature或cifar100_feature数据集配置。
+
+  notes:
+  - 所有数据集均为标准公开基准测试；除非需要离线执行，无需手动下载。
+  - 在使用迁移学习实验配置前，必须运行一次feature_extractor.py。
+credential_requirements:
+  mode: optional
+  items:
+  - name: Weights & Biases (wandb)
+    notes: experiment.py在启动时调用wandb.init()。执行前设置WANDB_API_KEY或运行`wandb login`。可通过设置WANDB_MODE=offline启用离线模式。
+
+  notes:
+  - 无需付费数据凭证。wandb是唯一的外部服务。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: 足以在MNIST或CIFAR-10上运行单实验quick_eval。
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 2
+    gpu_vram_gb: 24
+    notes: 两块GPU允许run_exp.py并行化执行实验套件。需要在6–24小时时间范围内完整复现所有四组消融图表。
+
+environment:
+  python: '3.8'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  framework: jax
+  key_packages:
+  - jax==0.5.2
+  - jaxlib==0.5.1
+  - jax-cuda12-pjrt==0.5.1
+  - jax-cuda12-plugin==0.5.1
+  - flax==0.10.4
+  - optax
+  - dp-accounting==0.4.4
+  - transformers==4.49.0
+  - datasets==2.21.0
+  - ml_collections==1.0.0
+  - wandb
+  - chex==0.1.89
+  - torch
+  - matplotlib==3.10.1
+  - jaxpruner @ git+https://github.com/google-research/jaxpruner@e7af4e4e4748782d6bc433249e359f74ffb845ba
+  notes:
+  - 本基准测试使用JAX/Flax作为主要框架，而非PyTorch。PyTorch仅在requirements.txt中用于DataLoader工具类。
+  - CUDA版本由jax-cuda12-* wheel决定；安装JAX CUDA包优先于其他依赖项。
+  - 完整固定依赖集请参见requirements.txt。
+  - 需要Python 3.8；使用的JAX版本（0.5.2）不支持3.12+。
+risk_flags:
+- jax_cuda_wheel_version_sensitivity
+- wandb_required_at_runtime
+- feature_extractor_prereq
+- no_runtime_execution_in_packaging_pass
+risk_notes:
+- JAX CUDA wheel对版本高度敏感。先安装jax-cuda12-pjrt和jax-cuda12-plugin再安装jaxlib，以避免ABI不匹配。
+
+- experiment.py无条件调用wandb.init()；无API key时设置WANDB_MODE=offline以防止硬性失败。
+
+- 使用cifar10_feature或cifar100_feature数据集配置的实验若未先运行feature_extractor.py将失败。
+
+- 本打包过程未执行任何基准测试运行。metric_contract.json中的所有指标值均为临时值，需运行时验证后方可信赖。请勿仅凭README信赖指标表面。
+
+recommended_when: >
+  在需要形式化(ε,δ)隐私保证的图像分类任务（CIFAR-10、CIFAR-100、MNIST）上评估认证机器遗忘方法时使用本基准测试。适合比较PABI、DP-SGD、收缩系数和输出扰动基线之间的遗忘效率（达到验证准确率阈值所需轮数）。最适合需要适用于非凸深度网络的无假设认证遗忘的研究人员。
+
+not_recommended_when: >
+  若仅需标准图像分类训练而无需隐私或遗忘约束，请勿使用。本基准测试不适用于需要PyTorch原生训练循环的场景——核心实现仅支持JAX/Flax。若GPU显存低于12GB或无法建立JAX CUDA环境，请勿使用。不适用于大规模语言模型遗忘或概念遗忘任务，这些任务需要不同的方法体系。
+
+paper:
+  title: Certified Unlearning for Neural Networks
+  authors:
+  - Anastasia Koloskova
+  - Youssef Allouah
+  - Animesh Jha
+  - Rachid Guerraoui
+  - Sanmi Koyejo
+  affiliations:
+  - Stanford University
+  - EPFL
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2506.06985
+  bibtex_key: koloskova2025certified
+  notes: Koloskova和Allouah同等贡献。加拿大温哥华第42届国际机器学习会议论文集。PMLR 267。
+
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.070_certifiedunlearning.zip
+  archive_type: zip
+  local_dir_name: paper-70-CertifiedUnlearning
+  upstream_repo: https://github.com/stair-lab/certified-unlearning-neural-networks-icml-2025
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.070_certifiedunlearning.zip
+  sha256: 124e11876363b6ccc532c5d235697f96975e0e2883506083a5f9391b63cf1ab7
+  size_bytes: 13612960
+commercial:
+  annual_fee: null
+display:
+  palette_seed: silver-plum-unlearn
+  art_style: privacy-lab
+  accent_priority: high
+image_path: ../image/070_aisb.t3.070_certifiedunlearning.jpg

package/AISB/catalog/aisb.t3.071_neuralmjd.yaml

@@ -0,0 +1,142 @@
+id: aisb.t3.071_neuralmjd
+name: Neural Non-Stationary Merton Jump Diffusion
+version: 0.1.0
+one_line: Neural network parameterization of non-stationary Merton jump diffusion
+  for time series prediction with explicit stochastic dynamics and abrupt jump modeling.
+task_description: 'This packaged benchmark covers jump-diffusion time-series prediction
+  using a neural parameterization of the non-stationary Merton jump diffusion (MJD)
+  model. The approach combines a time-inhomogeneous Itô diffusion to capture continuous
+  non-stationary stochastic dynamics with a time-inhomogeneous compound Poisson process
+  to model abrupt jumps. Parameters of the corresponding stochastic differential equations
+  (SDEs) are predicted by a neural network conditioned on historical data and contextual
+  information. The task evaluates forecasting fidelity on non-stationary data containing
+  discontinuous changes, using MAE for deterministic accuracy and NLL for probabilistic
+  calibration.
+
+  '
+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: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mae
+- nll
+- jump_diffusion_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run a single packaged jump-diffusion forecasting evaluation route on
+    pre-trained checkpoint with SP500 data.
+  estimated_time: 1-2h
+- id: full_train_eval
+  label: Full Train + Eval
+  description: Run the complete Monte Carlo jump-diffusion training workflow including
+    model initialization, training loop with periodic evaluation, checkpointing, and
+    final test set evaluation.
+  estimated_time: 6-18h
+dataset_download:
+  primary_method: curl
+  sources:
+  - name: SP500 Historical Stock Data
+    url: https://www.kaggle.com/api/v1/datasets/download/camnugent/sandp500
+    type: kaggle
+    format: zip
+    local_path: ./data/sandp500
+  notes:
+  - SP500 dataset requires Kaggle API credentials for download
+  - Demo notebook supports synthetic data generation without external downloads
+  - Configuration files located in config/sp500/mjd/
+credential_requirements:
+  mode: kaggle_api
+  items:
+  - name: Kaggle API Token
+    description: Required for SP500 dataset download from Kaggle
+    optional: true
+  notes:
+  - Kaggle credentials only needed for real-world SP500 data; synthetic data available
+    via demo notebook without credentials
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    network: false
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    network: true
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.2
+  flash_attn: null
+  key_packages:
+  - torch>=2.0.0
+  - pytorch-geometric
+  - pyyaml
+  - numpy
+  - scipy
+  notes:
+  - Environment provisioned via setup/conda.yaml or setup/requirements.txt
+  - PyG (pytorch-geometric) required for data loaders
+  - Full dependency set in bundled requirements; see README for conda setup instructions
+risk_flags:
+- long_training_time
+- stochastic_variance
+risk_notes:
+- Monte Carlo sampling introduces stochastic variance; multiple seeds recommended
+  for robust evaluation
+- Training on SP500 with default hyperparameters may take 6-12 hours on single GPU
+- Metric values are provisional pending first trusted execution; stochastic nature
+  means results may vary across runs
+recommended_when: 'Use this benchmark when you need probabilistic time-series forecasting
+  that explicitly models abrupt discontinuous jumps and non-stationary dynamics. Ideal
+  for financial time series with sudden price movements, supply chain data with promotional
+  spikes, or any domain requiring interpretable stochastic process modeling alongside
+  neural network flexibility.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you only need simple deterministic
+  point forecasting without uncertainty quantification. Not recommended when jump
+  events are annotated and event-based temporal point process models would be more
+  appropriate. Avoid when GPU memory < 16GB and training throughput is a primary constraint.
+
+  '
+paper:
+  title: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction
+  authors:
+  - Yuanpei Gao (University of British Columbia, Vector Institute)
+  - Qi Yan (University of British Columbia, Vector Institute)
+  - Yan Leng (University of Texas at Austin)
+  - Renjie Liao (University of British Columbia, Vector Institute, Canada CIFAR AI
+    Chair)
+  venue: arXiv preprint arXiv:2506.04542
+  year: 2025
+  url: https://arxiv.org/abs/2506.04542
+  github: https://github.com/DSL-Lab/neural-MJD
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.071_neuralmjd.zip
+  archive_type: zip
+  local_dir_name: paper-71-NeuralMJD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.071_neuralmjd.zip
+  sha256: 0ea7d70c8c22946e148ccfdf01295d7608888a1423b5ae6391ad927a6224f2e0
+  size_bytes: 101203
+display:
+  palette_seed: indigo-copper-jump
+  art_style: stochastic-signal
+  accent_priority: high
+image_path: ../image/071_aisb.t3.071_neuralmjd.jpg

package/AISB/catalog/aisb.t3.071_neuralmjd.zh.yaml

@@ -0,0 +1,120 @@
+id: aisb.t3.071_neuralmjd
+name: 神经非平稳Merton跳跃扩散
+version: 0.1.0
+one_line: 用于时间序列预测的神经非平稳Merton跳跃扩散网络参数化方法，具备显式随机动力学建模和突变跳跃捕捉能力。
+task_description: '该标准化基准测试涵盖了基于神经参数化非平稳Merton跳跃扩散（MJD）模型的跳跃扩散时间序列预测任务。该方法结合了时变Itô扩散以捕捉连续非平稳随机动力学，以及时变复合泊松过程以建模突变跳跃。相应随机微分方程（SDEs）的参数由神经网络根据历史数据和上下文信息预测。该任务评估在包含不连续变化的非平稳数据上的预测保真度，使用MAE衡量确定性精度，使用NLL评估概率校准。
+
+'
+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: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- mae
+- nll
+- jump_diffusion_checkpoint
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 在预训练检查点上使用SP500数据运行单个标准化跳跃扩散预测评估流程。
+  estimated_time: 1-2h
+- id: full_train_eval
+  label: 完整训练+评估
+  description: 运行完整的蒙特卡洛跳跃扩散训练工作流程，包括模型初始化、周期性评估的训练循环、检查点保存以及最终测试集评估。
+  estimated_time: 6-18h
+dataset_download:
+  primary_method: curl
+  sources:
+  - name: SP500历史股票数据
+    url: https://www.kaggle.com/api/v1/datasets/download/camnugent/sandp500
+    type: kaggle
+    format: zip
+    local_path: ./data/sandp500
+  notes:
+  - SP500数据集需要Kaggle API凭证才能下载
+  - 演示笔记本支持合成数据生成，无需外部下载
+  - 配置文件位于config/sp500/mjd/目录
+credential_requirements:
+  mode: kaggle_api
+  items:
+  - name: Kaggle API令牌
+    description: 从Kaggle下载SP500数据集所需
+    optional: true
+  notes:
+  - Kaggle凭证仅在使用真实SP500数据时才需要；演示笔记本可通过合成数据使用，无需凭证
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+    network: false
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    network: true
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: 2.1.2
+  flash_attn: null
+  key_packages:
+  - torch>=2.0.0
+  - pytorch-geometric
+  - pyyaml
+  - numpy
+  - scipy
+  notes:
+  - 环境通过setup/conda.yaml或setup/requirements.txt配置
+  - PyG（pytorch-geometric）用于数据加载器
+  - 完整依赖集在打包的requirements文件中； conda环境设置说明请参阅README
+risk_flags:
+- long_training_time
+- stochastic_variance
+risk_notes:
+- 蒙特卡洛采样引入随机方差；建议使用多个随机种子以获得稳健的评估结果
+- 使用默认超参数在SP500上训练在单个GPU上可能需要6-12小时
+- 指标值为临时值，待首次可信执行后确定；随机特性意味着不同运行之间结果可能有所差异
+recommended_when: '当需要明确建模突变跳跃和不连续变化的概率时间序列预测时使用此基准。适用于具有突然价格波动的金融时间序列、具有促销尖峰的供应链数据，或任何需要将可解释的随机过程建模与神经网络灵活性相结合的场景。
+
+'
+not_recommended_when: '如果仅需简单的确定性点预测而不需要不确定性量化，请勿使用此基准。当跳跃事件已标注且基于事件的时序点过程模型更合适时不推荐使用。当GPU内存小于16GB且训练吞吐量是主要约束时应避免。
+
+'
+paper:
+  title: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction
+  authors:
+  - Yuanpei Gao (University of British Columbia, Vector Institute)
+  - Qi Yan (University of British Columbia, Vector Institute)
+  - Yan Leng (University of Texas at Austin)
+  - Renjie Liao (University of British Columbia, Vector Institute, Canada CIFAR AI
+    Chair)
+  venue: arXiv preprint arXiv:2506.04542
+  year: 2025
+  url: https://arxiv.org/abs/2506.04542
+  github: https://github.com/DSL-Lab/neural-MJD
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.071_neuralmjd.zip
+  archive_type: zip
+  local_dir_name: paper-71-NeuralMJD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.071_neuralmjd.zip
+  sha256: 0ea7d70c8c22946e148ccfdf01295d7608888a1423b5ae6391ad927a6224f2e0
+  size_bytes: 101203
+display:
+  palette_seed: indigo-copper-jump
+  art_style: stochastic-signal
+  accent_priority: high
+image_path: ../image/071_aisb.t3.071_neuralmjd.jpg