onescience_run_manifest.yaml

onescience_manifest_version: "0.1"
resource_type: dataset

resource:
  id: OneScience/lagrangian
  name: lagrangian
  domain: cfd
  domain_tags: [cfd, particle_simulation, lagrangian_dynamics]
  task: lagrangian_particle_simulation_dataset
  task_tags: [train_data, eval_data, inference_input, dataset_validation]
  modalities: [particle_trajectory]
  input_formats: [tfrecord]
  output_formats: [tfrecord]
  summary: DeepMind Lagrangian Water 子数据集标准数据包，包含 train、valid、test 三个 TFRecord split 和 metadata.json。

platform_resource:
  primary:
    platform: modelscope
    repo_id: OneScience/lagrangian
    repo_type: dataset
    url: https://modelscope.cn/datasets/OneScience/lagrangian
    revision: main
    readme_path: README.md
    manifest_path: onescience_run_manifest.yaml
  mirrors: []
  access:
    visibility: public
    license: unknown

website_integration:
  enabled: true
  click_target:
    platform: modelscope
    resource_url: https://modelscope.cn/datasets/OneScience/lagrangian
  llm_handoff:
    readme_required: true
    manifest_required: true
    download_readme_first: true
    resolve_related_models: true
    default_run_goal: dataset_validation
    cwd_note: 如果使用 modelscope download --cache_dir 下载数据集，请先 cd 到实际下载后的数据集仓库根目录；模型侧将 ONESCIENCE_LAGRANGIAN_DATA_DIR 指向本仓库 data/Water。

runtime:
  enabled: true
  onescience_domain: cfd
  min_onescience_version: null
  supported_execution: [local_dataset_validation, model_training_input, model_inference_input]
  environment:
    exported_env:
      ONESCIENCE_LAGRANGIAN_DATA_DIR: <dataset_repo_root>/data/Water
  dependencies:
    python: ">=3.10"
    python_packages: [tensorflow, numpy, pyyaml]

onescience:
  repo: https://gitee.com/onescience-ai/onescience
  official_links:
    gitee:
      doc: https://gitee.com/onescience-ai/onescience-doc
      onescience: https://gitee.com/onescience-ai/onescience
      skills: https://gitee.com/onescience-ai/oneskills
    github:
      doc: https://github.com/onescience-ai/OneScience-doc
      onescience: https://github.com/onescience-ai/OneScience
      skills: https://github.com/onescience-ai/oneskills
  install:
    required_by_default: false
    command: bash install.sh cfd
  source_paths:
    - onescience/src/onescience/datapipes/cfd/deepmind_lagrangian.py
  compatibility:
    examples_path: onescience/examples/cfd/Lagrangian_MGN
    status: examples_compatible
  datapipe: onescience.datapipes.cfd.DeepMindLagrangianDatapipe

runtime_package:
  kind: standard_runtime_package
  package_root: .
  standard_layout:
    workdir: .
    data_dir: data/Water
    metadata_dir: metadata
    output_dir: validation_outputs
  apply_policy:
    mode: direct_use
    target: session_workdir
    overwrite: false
    protect_installed_onescience: true
  entry_files:
    - data/Water/metadata.json
    - data/Water/train.tfrecord
    - data/Water/valid.tfrecord
    - data/Water/test.tfrecord
    - metadata/lagrangian_water_schema.json
    - scripts/validate_lagrangian_dataset.py
    - files_sha256.jsonl
  entrypoints:
    preflight: scripts/validate_lagrangian_dataset.py
    validate: scripts/validate_lagrangian_dataset.py
    inference: null
    train: null
    finetune: null
    evaluate: null
    visualize: null
    deploy: null

files:
  model_files: []
  dataset_files:
    - id: water_tfrecord_splits
      path: data/Water/*.tfrecord
      role: train_eval_inference_data
      description_zh: Water 子数据集的 train、valid、test TFRecord 序列文件。
      required: true
      required_for: [dataset_validation, train, inference, evaluate]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: data/Water
    - id: water_metadata
      path: data/Water/metadata.json
      role: schema_and_normalization_stats
      description_zh: Water 数据的维度、时间步长、边界、半径、速度和加速度归一化统计。
      required: true
      required_for: [dataset_validation, train, inference, evaluate]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: data/Water/metadata.json
    - id: integrity_inventory
      path: files_sha256.jsonl
      role: file_size_sha256_inventory
      description_zh: 整理后数据文件大小和 SHA256 清单。
      required: true
      required_for: [dataset_validation]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: files_sha256.jsonl
    - id: integrity_summary
      path: data_integrity_summary.json
      role: source_to_standardized_integrity_summary
      description_zh: 原始数据与整理后数据的文件名、大小、SHA256 对照摘要。
      required: true
      required_for: [dataset_validation]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: data_integrity_summary.json
  config_files:
    - id: schema
      path: metadata/lagrangian_water_schema.json
      role: dataset_schema
      description_zh: TFRecord SequenceExample 字段、dtype、shape 和 split 数量说明。
      required: true
      required_for: [dataset_validation, train, inference]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: metadata/lagrangian_water_schema.json
  sample_files:
    - id: validation_script
      path: scripts/validate_lagrangian_dataset.py
      role: dataset_validation
      description_zh: 检查数据文件结构、metadata、TFRecord 首条样本 shape/dtype 和可选 SHA256。
      required: true
      required_for: [dataset_validation]
      source: {platform: modelscope, repo_id: OneScience/lagrangian, repo_type: dataset, download_method: command_ref, command_ref: commands.download.download_dataset}
      local_path: scripts/validate_lagrangian_dataset.py

dataset:
  format: [tfrecord_sequence_example]
  sample_unit: 一个 SequenceExample 对应一条粒子仿真序列。
  schema:
    path: metadata/lagrangian_water_schema.json
  splits:
    train: {path: data/Water/train.tfrecord, expected_sequences: 1000}
    validation: {path: data/Water/valid.tfrecord, expected_sequences: 30}
    test: {path: data/Water/test.tfrecord, expected_sequences: 30}
  metadata:
    path: data/Water/metadata.json
    dim: 2
    sequence_length: 1000
    dt: 0.0025
    default_connectivity_radius: 0.015
    normalization_stats: [vel_mean, vel_std, acc_mean, acc_std]

relations:
  required_datasets: []
  optional_datasets: []
  compatible_models:
    - id: OneScience/LagrangianMGN
      role: train_data
      required_for: [preflight, train, inference, evaluate, visualize]
      resource_ref:
        platform: modelscope
        repo_id: OneScience/LagrangianMGN
        repo_type: model
        url: https://modelscope.cn/models/OneScience/LagrangianMGN
        revision: main
        readme_path: README.md
        manifest_path: onescience_run_manifest.yaml
      expected_local_path: data/Water
      adapter: null

run_matrix:
  scenarios:
    - name: dataset_validation
      default: true
      capability: preflight
      description: 验证 Water TFRecord 数据、metadata schema、文件清单和可读性。
      required_datasets:
        - id: OneScience/lagrangian
          role: dataset_validation
          local_path: .
      required_model_files: []
      required_dataset_files: [data/Water/metadata.json, data/Water/train.tfrecord, data/Water/valid.tfrecord, data/Water/test.tfrecord, files_sha256.jsonl]
      preconditions:
        - 已下载 OneScience/lagrangian 数据集仓库并 cd 到数据集仓库根目录。
      command_refs: [commands.preflight.validate_dataset]
      expected_outputs: [dataset_validation_ok]

capabilities:
  inference: false
  train: false
  finetune: false
  evaluate: false
  visualize: false
  deploy: false
  dataset_validation: true

commands:
  download:
    - name: download_dataset
      target: dataset
      repo_id: OneScience/lagrangian
      local_path: <session_workdir>/lagrangian
      required: true
      required_for: [dataset_validation, train, inference, evaluate]
      command: modelscope download --dataset OneScience/lagrangian
  preflight:
    - name: validate_dataset
      description: 验证 Water 数据集文件结构、metadata、TFRecord 首条样本和清单。
      cwd: .
      command: python scripts/validate_lagrangian_dataset.py --dataset-root .
      required_files: [scripts/validate_lagrangian_dataset.py, data/Water/metadata.json, data/Water/train.tfrecord, data/Water/valid.tfrecord, data/Water/test.tfrecord, files_sha256.jsonl]
      required_env: []
      expected_outputs: [{path: stdout, type: text, contains: "Lagrangian Water dataset validation passed"}]
      success_criteria: [exit_code == 0]
  prepare: []
  inference: []
  train: []
  finetune: []
  evaluate: []
  visualize: []
  deploy: []

expected_outputs:
  - id: dataset_validation_ok
    path: stdout
    type: text
    description_zh: 数据集验证输出 Lagrangian Water dataset validation passed。

diagnostics:
  - symptom: "missing TFRecord split"
    cause: 数据集下载不完整或当前目录不是数据集仓库根目录。
    fix: 重新执行 modelscope download --dataset OneScience/lagrangian，并 cd 到下载后的数据集根目录。
  - symptom: "metadata dim or sequence_length mismatch"
    cause: 使用了非 Water 或非本标准包的数据目录。
    fix: 确认 ONESCIENCE_LAGRANGIAN_DATA_DIR 指向 OneScience/lagrangian 的 data/Water。
  - symptom: "TensorFlow first-record check skipped"
    cause: 当前环境缺少 TensorFlow。
    fix: 安装 TensorFlow 后重跑验证；文件大小和 SHA256 仍可先用于完整性判断。

domain_extension:
  cfd:
    dataset_family: DeepMind Learning to Simulate
    subset: Water
    simulation_type: lagrangian_particle_dynamics
    dimensionality: 2
    splits: [train, valid, test]