Datasets:
Contributing
DocuBench is intended to be reproducible benchmark infrastructure. Contributions should preserve that property.
Development Setup
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e ".[dev]"
Checks
Run these before opening a pull request:
docubench validate
docubench score
pytest
Result Submissions
Add new system outputs under results/<system_name>/. Each <doc_id>.json file should include a top-level data object. See docs/submissions.md.
If your system uses a prompt or run configuration, commit it under prompts/ alongside the existing baselines so the result set is reproducible and auditable.
Leaderboard Data
The Hugging Face Space in space/ renders space/leaderboard.json. After changing results or labels, refresh it so the leaderboard stays in sync:
docubench report
python3 space/build_data.py
tests/test_space.py fails if the committed space/leaderboard.json is stale.
Dataset Changes
Changes to documents, schemas, labels, or scoring behavior can change benchmark meaning. Include:
- the reason for the change
- affected document IDs
- before/after scoring impact if applicable
- source and license details for any new document
Scoring changes should be treated as benchmark-version changes.