Fix GAIA attachment downloads

#2

Summary

The current GAIA dataset provides attachment file_path values as repository-relative paths, but the Space currently checks them directly as local filesystem paths.

This change uses hf_hub_download() to resolve repository-relative attachment paths to files in the local Hugging Face cache. Existing absolute local paths continue to use the previous behavior.

Validation

Tested on a private duplicate of the current Space:

  • The application loaded all 20 questions successfully.
  • All 5 referenced attachments were downloaded and mapped.
  • /files/{task_id} returned HTTP 200 for:
    • PNG
    • MP3 (2 files)
    • Python
    • XLSX

The change is intentionally limited to attachment path resolution; unrelated loading and API behavior are unchanged.

Related reports

Reproduced on 31 Oct 2026 against the live public scoring API (no authentication or private dataset access): GET /questions returns all 20 tasks (HTTP 200), but every one of the five tasks with file_name returns HTTP 404 from the documented GET /files/{task_id} endpoint, with No file path associated with task_id. Affected attachment types: 1 PNG, 2 MP3, 1 Python file, 1 XLSX. This prevents legitimate agent implementations from solving the full 20-task set, independent of the learner's own Space. Your scoped path-resolution fix appears to address the precise failure; please review/merge when practical. No benchmark answers or ground truth are involved.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment