fieldbench/

corpus

A cross-domain, field-level benchmark for schema-driven document extraction (document → structured JSON). 1,441 documents across 10 categories with per-field ground truth, released so extraction-accuracy claims become falsifiable and comparable.

Code / scorer: https://github.com/fieldbench/fieldbench (pip install fieldbench) Full corpus + datasheet: https://github.com/fieldbench/corpus Datasheet: see DATASHEET.md in the corpus repo — read it before drawing… See the full description on the dataset page: https://huggingface.co/datasets/fieldbench/corpus.