MDPBench Leaderboard

Static, official leaderboard for MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios.

The deployed page needs no server or API key. Update leaderboard.json to publish a new verified result. The initial entries are the results in the repository's main-results table; source_name keeps the original table label where the displayed official release/API identifier is more specific. The initial board contains all 31 models from that table.

Before publishing an updated result table, synchronize the data and verify the result:

python tools/sync_leaderboard.py --write
python tools/sync_leaderboard.py --check

Deployment

This directory is the shared static-site source for both public deployments:

• Hugging Face: https://huggingface.co/spaces/Delores-Lin/MDPBench-leaderboard
• GitHub Pages: https://delores-lin.github.io/MultimodalOCR/ (after GitHub Pages is enabled for the repository)

The deploy-mdpbench-leaderboard GitHub Actions workflow publishes this same directory to both services on each main-branch update. GitHub Pages needs no credential. For Hugging Face, add an HF_TOKEN Actions secret with write access to Delores-Lin/MDPBench-leaderboard; the workflow skips that deployment until the secret is present.

Result policy

Public scores can be reproduced using the released evaluation set. Private scores are evaluated by the MDPBench maintainers; a model is ranked by private score only after that verification. See the benchmark README for the private-set submission procedure.