eval_752
  • English

      • Dataset Export Guide

      The dataset explorer and automation scripts share the same export pipeline: a FastAPI endpoint that streams a .eval752.zip archive. This guide summarizes how to retrieve packages, how filtering impacts the archive, and how the UI exposes the workflow.

      Endpoint Summary

      GET /datasets/{dataset_id}/export

      Query paramTypeDescription
      run_idstring (optional)Include run metadata, run_config.json, and results.jsonl for a completed run belonging to the dataset.
      section_idsstring (repeatable, optional)Limit the export to specific section IDs. Pass multiple section_ids pairs to include more than one section.

      Response: 200 OK with Content-Type: application/zip and Content-Disposition: attachment; filename=...eval752.zip.

      The archive always contains:

      • manifest.json — generation metadata, section list, and counts that honor filters.
      • meta.json — optional description/schema metadata.
      • sections/*.jsonl — one JSONL file per exported section.
      • assets/... — decoded binary assets referenced by exported items.
      • When run_id is provided: run_config.json, results.jsonl, and LightEval config (if available).

      Section Filtering

      • The backend accepts any number of section_ids parameters. IDs that do not belong to the dataset trigger 400 Bad Request to prevent accidental mismatches.
      • Filters apply to both dataset content and run artifacts. results.jsonl only includes items from the exported sections, so downstream tools never see dangling references.
      • When no section_ids are specified, all sections are exported.

      Example:

      curl -L -o dataset-sections.eval752.zip \
  "http://localhost:8000/datasets/$DATASET_ID/export?section_ids=$SECTION_A&section_ids=$SECTION_B"

      Run Selection

      Passing run_id instructs the service to embed the matching run summary, run_config.json, LightEval config (without API keys), and results.jsonl. The run must belong to the dataset; otherwise the API returns 400 with a descriptive error.

      curl -L -o run-packaged.eval752.zip \
  "http://localhost:8000/datasets/$DATASET_ID/export?run_id=$RUN_ID"

      Combine run_id with section_ids to focus on a subset of sections while keeping the run metrics for those items only.

      UI & UX Notes

      • The dataset list still exposes a one-click “Export” button for full packages.
      • The in-app Dataset Explorer now adds an Export view control that pulls the filtered sections a user is reviewing. Microcopy on the button clarifies that search keywords are for previewing only, while section filters define the export scope—reducing surprise for users who expect WYSIWYG downloads.
      • Run detail panels continue to provide the most direct “export with run results” entry point.

      Keep these cues aligned in future design work so that the mental model remains consistent: filters = structural subsets, runs = behavioral context.

      Troubleshooting

      SymptomLikely causeResolution
      400 Sections do not belong to this datasetA section_ids value came from another dataset or is stale.Refresh the dataset list and copy IDs from GET /datasets/{id} before retrying.
      400 Run belongs to a different datasetThe selected run references a separate dataset.Export from the Run detail view or verify the dataset ID in the URL.
      Empty results.jsonl even with run_idThe filtered sections excluded every item the run evaluated.Remove the section filter or include the relevant sections.