@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.050_causalpfn.zh.yaml

@@ -0,0 +1,190 @@
+id: aisb.t3.050_causalpfn
+name: 'CausalPFN：通过上下文学习实现摊销式因果效应估计'
+version: 0.1.0
+one_line: 基于Transformer的摊销式因果效应估计，采用PFN风格的上下文学习，具备校准的不确定性量化。
+task_description: 'CausalPFN是一个Transformer模型，通过在一个包含大量模拟数据生成过程（DGP）的大型库上进行训练来实现摊销式因果效应估计，这些数据生成过程满足可忽略性假设。面对新的观测数据集时，它通过上下文学习直接推断因果效应，无需任务特定的调优、微调或超参数优化。该模型同时估计条件平均处理效应（CATE）和平均处理效应（ATE），并通过基于原则的不确定性量化框架产生有限样本校准的可信区间。摊销式方法将计算负担从推理阶段转移到预训练阶段，使得在新数据集上的评估仅需快速的前向传播即可完成。'
+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: 2-8h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- ihdp_pehe
+- ihdp_ate_rel_err
+- causal_predictions
+- bootstrap_report
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 在单个支持的数据集（如IHDP）上使用默认设置运行CausalPFN评估。适用于验证安装和基本功能。
+- id: full_benchmark
+  label: 完整基准
+  description: 在IHDP、ACIC和Lalonde数据集上运行完整的摊销式因果效应基准测试，包含Bootstrap集成聚合和温度缩放。
+dataset_download:
+  primary_method: pypi
+  sources:
+  - name: PyPI
+    url: https://pypi.org/project/causalpfn/
+    archive_type: null
+    local_dir_name: null
+  - name: HuggingFace Hub
+    url: https://huggingface.co/vdblm/causalpfn
+    archive_type: null
+    local_dir_name: null
+  notes:
+  - 模型权重在运行时从HuggingFace Hub下载。
+  - 安装方式：pip install causalpfn
+  - 可选的开发依赖：pip install causalpfn[dev]
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - HuggingFace Hub访问需要网络连接以下载模型。
+  - 可选的HF_TOKEN用于受限或私有模型的访问。
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes: 通过causalpfn Python API的torch.device('cpu')支持仅CPU执行。
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+    notes: GPU加速推荐用于在大型数据集上实现更快的推理。
+environment:
+  python: '3.10'
+  cuda: '11.8'
+  pytorch: '2.3'
+  flash_attn: null
+  key_packages:
+  - causalpfn>=1.0.0
+  - torch>=2.3.0
+  - numpy>=1.24.0
+  - scikit-learn>=1.3.0
+  - faiss-cpu==1.9.0
+  - huggingface_hub>=0.20.0
+  - tqdm>=4.65.0
+  notes:
+  - 仅CPU执行适用于最低资源配置。
+  - 通过CUDA实现GPU加速以提升推理速度。
+  - 完整依赖集可在捆绑的requirements.txt和pyproject.toml中找到。
+risk_flags:
+- external_model_download
+- bootstrap_variance
+risk_notes:
+- 模型权重在运行时从HuggingFace Hub获取；请确保网络访问。
+- Bootstrap集成在不同随机种子间引入方差；使用提供的bootstrap_report进行诊断。
+recommended_when: '当因果推理任务需要易于本地运行的方案、清晰的效果估计指标（PEHE、ATE相对误差）和不确定性量化时，可使用此基准。是 uplift建模、从观测数据估计处理效应以及与专业因果估计器进行比较的理想选择。当需要无需手动调优的开箱即用估计器时，推荐使用。'
+not_recommended_when: '如果需要大型多模态基准、原始语言生成任务或图像/视频理解，请勿使用此基准。当处理分配不可忽略或数据中存在未观测到的混杂因素时，不适用。'
+paper:
+  title: 'CausalPFN: Amortized Causal Effect Estimation via In-Context Learning'
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://arxiv.org/abs/2506.07918
+  authors:
+  - Vahid Balazadeh
+  - Hamidreza Kamkari
+  - Valentin Thomas
+  - Benson Li
+  - Junwei Ma
+  - Jesse C. Cresswell
+  - Rahul G. Krishnan
+  institutions:
+  - University of Toronto
+  - Vector Institute
+  - Layer 6 AI
+display:
+  palette_seed: mint-charcoal-causal
+  art_style: causal-analytics
+  accent_priority: medium
+image_path: ../image/050_aisb.t3.050_causalpfn.jpg
+capability_tags:
+- causal_inference
+- in_context_learning
+- tabular_ml
+- uncertainty_estimation
+- uplift_modeling
+- treatment_effect_estimation
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+code_anchors:
+  evaluation:
+  - path: eval.py
+    description: Main evaluation script for IHDP benchmark with multi-seed bootstrap
+      ensemble
+  - path: src/causalpfn/evaluation.py
+    description: Core evaluation metrics including PEHE and calibration curve calculations
+  estimators:
+  - path: src/causalpfn/causal_estimator.py
+    description: CATEEstimator and ATEEstimator classes with confidence interval estimation
+  benchmarks:
+  - path: benchmarks/ihdp.py
+    description: IHDP benchmark implementation
+  - path: benchmarks/acic2016.py
+    description: ACIC 2016 benchmark implementation
+  - path: benchmarks/lalonde.py
+    description: Lalonde benchmark implementation
+  - path: benchmarks/hillstrom.py
+    description: Hillstrom uplift modeling benchmark
+metric_contract:
+  ihdp_pehe:
+    origin_path: eval.py
+    source_ref: calculate_pehe
+    evaluation_protocol:
+      code_paths:
+      - eval.py
+      - src/causalpfn/evaluation.py
+    status: code_backed
+    provisional: false
+  ihdp_ate_rel_err:
+    origin_path: eval.py
+    source_ref: calculate_ate_rel_error
+    evaluation_protocol:
+      code_paths:
+      - eval.py
+      - src/causalpfn/evaluation.py
+    status: code_backed
+    provisional: false
+evaluation_details:
+  bootstrap_ensemble:
+    seeds: 2
+    subsamples: 3
+    subsample_fraction: 0.92
+    temperature_values:
+    - 0.3
+    - 0.5
+    - 0.7
+    - 0.9
+    - 2.0
+    - 4.0
+    - 8.0
+    total_predictions_per_realization: 42
+  s_learner:
+    regressor: Ridge
+    alpha: 0.001
+    mu_features: true
+  nearest_neighbors: 85
+commercial:
+  annual_fee: null
+  notes: Apache 2.0 License
+download:
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.050_causalpfn.zip
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.050_causalpfn.zip
+  archive_type: zip
+  sha256: d7ba4ba97509f90f00b4f9953b10d76fc5bc56ea116a7141a74a11990ee99a99
+  size_bytes: 48053403

package/AISB/catalog/aisb.t3.051_flashtp.yaml

@@ -0,0 +1,169 @@
+schema_version: 1
+id: aisb.t3.051_flashtp
+name: 'FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic
+  Potentials'
+version: 0.1.0
+one_line: 'Benchmark and optimize fused, sparsity-aware CUDA tensor-product kernels
+  that accelerate equivariant machine-learning interatomic potentials (MLIPs), measuring
+  forward, backward, and double-backward kernel latency in milliseconds.
+
+  '
+task_description: 'This benchmark evaluates CUDA kernel performance for the Tensor-Product
+  layer used in equivariant machine-learning interatomic potentials (MLIPs). The Tensor-Product
+  layer is the dominant computational bottleneck in equivariant MLIP models such as
+  SevenNet, accounting for up to 89% of inference time and 75% of training time. FlashTP
+  addresses this by fusing all kernels within the Tensor-Product layer and its subsequent
+  reduce layer into a single kernel, exploiting the high sparsity (71–86%) of the
+  Clebsch-Gordan coefficient matrix, and introducing path-aggregated execution to
+  increase input reuse.
+
+  The evaluation route is self-contained: run `example/benchmark_timing.py` with IR
+  configs from `ir_config/` (NequIP, MACE, SevenNet configurations are bundled). The
+  script times forward, backward, and double-backward passes using CUDA events (50
+  warmup iterations, 200 timed iterations) and reports per-iteration latency in milliseconds.
+  A simpler smoke test is available via `example/test.sh`. No external datasets, credentials,
+  or evaluation servers are required. Custom CUDA extensions must be compiled from
+  source (`pip install . --no-build-isolation`), which takes up to 10 minutes. By
+  default, kernels are compiled for NVIDIA A100 (sm_80) and H100 (sm_90); set `CUDA_ARCH_LIST`
+  to target other architectures.
+
+  The optimization task is to improve kernel throughput (lower forward_ms, backward_ms,
+  double_backward_ms) without violating numerical correctness against the reference
+  e3nn TensorProduct implementation.
+
+  '
+capability_tags:
+- research_code_optimization
+- cuda_kernels
+- scientific_ml
+- interatomic_potentials
+- systems_efficiency
+- gpu_kernel_fusion
+- sparse_computation
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 2-6h
+difficulty: hard
+data_access: public
+primary_outputs:
+- forward_ms
+- backward_ms
+- double_backward_ms
+- kernel_benchmark
+launch_profiles:
+- id: microbenchmark
+  label: Microbenchmark
+  description: 'Run example/benchmark_timing.py on a single IR config to measure forward,
+    backward, and double-backward kernel latency. Fastest route; completes in minutes
+    per config.
+
+    '
+- id: full_eval
+  label: Full Eval
+  description: 'Run benchmarks across all bundled IR configs (NequIP, MACE, SevenNet)
+    with multiple edge counts and data types to produce a comprehensive kernel performance
+    profile.
+
+    '
+- id: smoke_test
+  label: Smoke Test
+  description: 'Run example/test.sh to verify that the CUDA extension compiles and
+    the FlashTP kernel produces correct output on one NequIP config (4_32_3_p_sc,
+    32768 edges, fp32).
+
+    '
+dataset_download:
+  primary_method: none
+  sources: []
+  notes:
+  - No external dataset download is required. All IR configs for microbenchmarking
+    are bundled in ir_config/.
+  - Edge graphs are synthetically generated at runtime by flashTP_e3nn.utils.fixed_generate_edgepair.
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 80
+environment:
+  python: '3.10'
+  cuda: '12'
+  pytorch: 2.4.1
+  key_packages:
+  - e3nn==0.6.6
+  - torch_scatter
+  notes:
+  - The README explicitly requires CUDA Toolkit 12 or higher.
+  - Python 3.11 is mentioned as non-strict; 3.10+ should work.
+  - Compiling custom CUDA kernels takes up to 10 minutes during pip install.
+  - Default CUDA_ARCH_LIST targets sm_80 (A100) and sm_90 (H100). Set CUDA_ARCH_LIST
+    to match your GPU.
+  - torch_scatter is required for accurate e3nn reference benchmarks.
+  - See requirements.txt and pyproject.toml for the full dependency set.
+risk_flags:
+- custom_cuda_compilation
+- gpu_architecture_dependency
+risk_notes:
+- CUDA extensions must compile from source; compilation failure on unsupported architectures
+  is a common failure mode.
+- Default build targets A100/H100 only. Other GPUs require setting CUDA_ARCH_LIST
+  to the correct compute capability.
+- No benchmark execution was performed in the packaging pass; metric values should
+  be verified at runtime.
+- The e3nn reference implementation is used for numerical correctness checks; mismatched
+  e3nn versions may cause subtle differences.
+recommended_when: 'Use this benchmark when you want a self-contained CUDA-kernel optimization
+  task focused on equivariant tensor-product operations in atomistic ML workloads.
+  Ideal for evaluating kernel fusion, sparsity exploitation, and GPU memory optimization
+  strategies on a well-defined microbenchmark with forward/backward/double-backward
+  phases.
+
+  '
+not_recommended_when: 'Do not use this if you cannot compile custom CUDA extensions,
+  lack an NVIDIA GPU with compute capability 8.0+, or need a high-level application
+  benchmark with end-to-end MD simulation evaluation. Not suitable for CPU-only environments.
+
+  '
+paper:
+  title: 'FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic
+    Potentials'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/lee25l.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.051_flashtp.zip
+  archive_type: zip
+  local_dir_name: paper-51-FlashTP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.051_flashtp.zip
+  sha256: c347ddfa54284f9900fb0629240f73461930b7d0eaa83a68b1bfce61cf6d5fb6
+  size_bytes: 148422
+commercial:
+  annual_fee: null
+display:
+  palette_seed: graphite-amber-tensor
+  art_style: hpc-kernel
+  accent_priority: high
+image_path: ../image/051_aisb.t3.051_flashtp.jpg

package/AISB/catalog/aisb.t3.051_flashtp.zh.yaml

@@ -0,0 +1,124 @@
+schema_version: 1
+id: aisb.t3.051_flashtp
+name: FlashTP：用于机器学习原子间势能融合的稀疏感知张量积
+version: 0.1.0
+one_line: >
+  对融合的稀疏感知CUDA张量积内核进行基准测试和优化，加速等变机器学习原子间势能（MLIP），以毫秒为单位测量前向、反向和双反向内核延迟。
+task_description: >
+  本基准测试评估等变机器学习原子间势能（MLIP）中使用的张量积层的CUDA内核性能。张量积层是等变MLIP模型（如SevenNet）中的主要计算瓶颈，占推理时间的89%和训练时间的75%。FlashTP通过将张量积层及其后续归约层中的所有内核融合为单一内核、利用Clebsch-Gordan系数矩阵的高稀疏性（71-86%），并引入路径聚合执行以增加输入重用来解决这一问题。
+
+  评估流程完全自包含：使用ir_config/中的IR配置（NequIP、MACE、SevenNet配置已捆绑）运行example/benchmark_timing.py。该脚本使用CUDA事件对前向、反向和双反向传播进行计时（50次预热迭代，200次计时迭代），并以毫秒为单位报告每次迭代的延迟。可以通过example/test.sh进行更简单的冒烟测试。无需外部数据集、凭证或评估服务器。必须从源码编译自定义CUDA扩展（pip install . --no-build-isolation），这可能需要长达10分钟。默认情况下，内核针对NVIDIA A100（sm_80）和H100（sm_90）编译；设置CUDA_ARCH_LIST以针对其他架构。
+
+  优化任务是提高内核吞吐量（降低forward_ms、backward_ms、double_backward_ms），同时不违反与参考e3nn TensorProduct实现的数值正确性。
+capability_tags:
+- research_code_optimization
+- cuda_kernels
+- scientific_ml
+- interatomic_potentials
+- systems_efficiency
+- gpu_kernel_fusion
+- sparse_computation
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+cost_band: high
+time_band: 2-6h
+difficulty: hard
+data_access: public
+primary_outputs:
+- forward_ms
+- backward_ms
+- double_backward_ms
+- kernel_benchmark
+launch_profiles:
+- id: microbenchmark
+  label: 微观基准
+  description: >
+    在单个IR配置上运行example/benchmark_timing.py，测量前向、反向和双反向内核延迟。最快的方式；每个配置几分钟内完成。
+- id: full_eval
+  label: 完整评估
+  description: >
+    跨所有捆绑的IR配置（NequIP、MACE、SevenNet）运行基准测试，包含多个边数和数据类型，以生成全面的内核性能配置文件。
+- id: smoke_test
+  label: 冒烟测试
+  description: >
+    运行example/test.sh验证CUDA扩展编译成功，并在一种NequIP配置（4_32_3_p_sc，32768条边，fp32）上确认FlashTP内核产生正确输出。
+dataset_download:
+  primary_method: none
+  sources: []
+  notes:
+  - 无需下载外部数据集。所有用于微观基准测试的IR配置都捆绑在ir_config/中。
+  - 边图在运行时由flashTP_e3nn.utils.fixed_generate_edgepair合成生成。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 24
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 120
+    gpu_count: 1
+    gpu_vram_gb: 80
+environment:
+  python: '3.10'
+  cuda: '12'
+  pytorch: 2.4.1
+  key_packages:
+  - e3nn==0.6.6
+  - torch_scatter
+  notes:
+  - README明确要求CUDA Toolkit 12或更高版本。
+  - Python 3.11被提及为非严格要求；3.10+应该可以工作。
+  - 在pip install期间编译自定义CUDA内核最多需要10分钟。
+  - 默认CUDA_ARCH_LIST针对sm_80（A100）和sm_90（H100）。设置CUDA_ARCH_LIST以匹配您的GPU。
+  - torch_scatter是精确e3nn参考基准测试所必需的。
+  - 完整的依赖项列表请参见requirements.txt和pyproject.toml。
+risk_flags:
+- custom_cuda_compilation
+- gpu_architecture_dependency
+risk_notes:
+  - CUDA扩展必须从源码编译；在不支持的架构上编译失败是常见的失败模式。
+  - 默认构建仅针对A100/H100。其他GPU需要将CUDA_ARCH_LIST设置为正确的计算能力。
+  - 打包过程中未执行基准测试运行；指标值应在运行时验证。
+  - e3nn参考实现用于数值正确性检查；不匹配的e3nn版本可能导致细微差异。
+recommended_when: >
+  当您需要一个专注于原子级ML工作负载中等变张量积操作的独立CUDA内核优化任务时使用此基准测试。非常适合评估内核融合、稀疏性利用和GPU内存优化策略，使用具有前向/反向/双反向阶段的定义明确的微观基准。
+not_recommended_when: >
+  如果您无法编译自定义CUDA扩展、缺少计算能力8.0+的NVIDIA GPU，或需要带有端到端MD模拟评估的高级应用程序基准测试，请勿使用此基准测试。不适合纯CPU环境。
+paper:
+  title: 'FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic
+    Potentials'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/lee25l.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.051_flashtp.zip
+  archive_type: zip
+  local_dir_name: paper-51-FlashTP
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.051_flashtp.zip
+  sha256: c347ddfa54284f9900fb0629240f73461930b7d0eaa83a68b1bfce61cf6d5fb6
+  size_bytes: 148422
+commercial:
+  annual_fee: null
+display:
+  palette_seed: graphite-amber-tensor
+  art_style: hpc-kernel
+  accent_priority: high
+image_path: ../image/051_aisb.t3.051_flashtp.jpg

package/AISB/catalog/aisb.t3.052_nsdiff.yaml

@@ -0,0 +1,155 @@
+id: aisb.t3.052_nsdiff
+name: Non-stationary Diffusion For Probabilistic Time Series Forecasting
+version: 0.1.0
+one_line: Diffusion-based probabilistic forecasting framework for non-stationary time-series
+  dynamics, evaluated via CRPS on ETTh benchmarks.
+task_description: 'NsDiff is a diffusion-based theoretical framework for probabilistic
+  forecasting designed specifically for non-stationary scenarios. The packaged benchmark
+  covers the training and evaluation pipeline for this method, including pretraining
+  scripts, main experiment entry point at ./src/experiments/NsDiff.py, and configuration
+  files. However, the current local snapshot does not preserve the full executable
+  source tree referenced in the README; the code-bearing roots (configs/, notebooks/)
+  provide partial anchors only. The benchmark requires upstream restoration before
+  full benchmark execution can be performed. Datasets are downloaded automatically
+  during execution. Evaluation uses CRPS (Continuous Ranked Probability Score) as
+  the primary metric.
+
+  '
+task_mode: analysis_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: restore_needed
+support_level: recovery
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- crps
+launch_profiles:
+- id: analysis_only
+  label: Analysis Only
+  description: Inspect preserved notebook at notebooks/uncertainty_variation.ipynb,
+    configuration at configs/nsdiff.yml, and bundled documentation for the NsDiff
+    route.
+- id: restore_first
+  label: Restore First
+  description: Restore the missing source tree referenced in the README before attempting
+    benchmark execution. Requires recovering ./src/experiments/NsDiff.py and ./scripts/
+    directory from upstream.
+dataset_download:
+  primary_method: automatic
+  sources:
+  - The README states datasets are downloaded automatically during execution.
+  - Referenced benchmarks include ETTh1, ETTh2, ETTm1, ETTm2, and other ETT dataset
+    variants.
+  notes:
+  - Treat dataset setup as part of upstream restoration; no bundled download scripts
+    preserved in current snapshot.
+  - Execution commands reference dataset_type parameter (e.g., --dataset_type="ETTh1").
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 200
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '11.8'
+  pytorch: '2.0'
+  flash_attn: null
+  key_packages:
+  - See ./requirements.txt for full dependency set
+  - Core dependencies inferred from configs/nsdiff.yml testing configuration
+  notes:
+  - The current local package does not preserve executable setup files or runnable
+    source code beyond partial configs/ and notebooks/ roots.
+  - Environment hints are provisional pending upstream restoration.
+  - Install command: pip install -r ./requirements.txt
+risk_flags:
+- source_snapshot_incomplete
+- route_caveat
+- crps_metric_blocked
+risk_notes:
+- The current snapshot does not expose executable code files that can anchor the CRPS
+  metric.
+- No benchmark execution was performed in this packaging pass; treat all metric claims
+  as provisional.
+- The configs/nsdiff.yml provides sampling and testing parameters (batch_size=256,
+  sampling_size=1000, n_z_samples=200, n_bins=10) but no executable evaluation anchor.
+recommended_when: Use this entry when you want paper-backed planning, upstream restoration
+  guidance, or exploratory analysis of the preserved notebook and configuration files
+  for the NsDiff diffusion forecasting route.
+not_recommended_when: Do not treat this local package as a ready-to-run diffusion
+  benchmark. The README-referenced source tree (src/experiments/, scripts/) is not
+  preserved locally. Full benchmark execution requires upstream restoration of the
+  executable source tree before CRPS evaluation can be performed.
+paper:
+  title: Non-stationary Diffusion For Probabilistic Time Series Forecasting
+  venue: ICML 2025 Spotlight
+  year: 2025
+  url: https://arxiv.org/abs/2505.08613
+  authors:
+  - Weiwei Ye (equal contribution)
+  - Zhuopeng Xu (equal contribution)
+  - Ning Gui (corresponding author)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.052_nsdiff.zip
+  archive_type: zip
+  local_dir_name: paper-52-NsDiff
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.052_nsdiff.zip
+  sha256: cc513276d021fd69cb749c3d843a64dc99b18428e9d1d66f040d07d62bed139a
+  size_bytes: 435952
+display:
+  palette_seed: navy-mint-diffusion
+  art_style: probabilistic-signal
+  accent_priority: high
+image_path: ../image/052_aisb.t3.052_nsdiff.jpg
+execution_evidence:
+  pretrain_script: bash ./scripts/pretrain_F/ETTh1.sh
+  main_entry: python3 ./src/experiments/NsDiff.py
+  example_args:
+  - --dataset_type=ETTh1
+  - --device=cuda:0
+  - --batch_size=32
+  - --horizon=1
+  - --pred_len=192
+  - --windows=168
+  - --load_pretrain=True
+  notes:
+  - PYTHONPATH must be set to project root during execution.
+  - CUDA_DEVICE_ORDER=PCI_BUS_ID recommended for multi-GPU setups.
+  - Seeds parameter accepts list format: --seeds='[1232132, 3]'
+  - --load_pretrain=False option available for non-pretrained runs.
+config_anchors:
+- path: configs/nsdiff.yml
+  role: Primary configuration file
+  contains:
+  - sampling: batch_size=256, sampling_size=1000, last_only=True
+  - testing: batch_size=256, n_z_samples=200, n_bins=10, minisample=1
+  - coverage and NLL tracking parameters
+- path: notebooks/uncertainty_variation.ipynb
+  role: Analysis notebook anchor
+- path: configs/csbi.yaml
+  role: Additional config (not primary)
+- path: configs/tmdm.yml
+  role: Additional config (not primary)
+code_audit_status: blocked
+metric_blocking_reason: No executable code anchor for CRPS evaluation in current snapshot.
+  The configs/nsdiff.yml contains testing parameters but the actual evaluation code
+  path is not preserved.