📄 DocuBench Leaderboard

50 hard, real-world documents · schema-guided structured extraction · v0.1.0

Each task pairs a source document with a JSON Schema and a hand-verified JSON label. Systems are scored on field-level accuracy with order-independent array matching (metric: macro_average_field_accuracy). Every number here is produced by the same public scorer against the same labels — these are baseline submissions, not a closed leaderboard.

Code, data, scorer, and committed prompts: github.com/DocuPipe/docubench