davanstrien/

glam-extraction-bench-poc

A proof-of-concept for an open GLAM document-extraction benchmark (index cards + registration forms). Each item = a card/form image + a target JSON Schema + a target output JSON.

Silver labels, not a finished benchmark. Outputs are NuExtract-3 zero-shot, reviewed/corrected by one LLM reviewer agent — not human-verified. NuExtract is therefore advantaged. Treat scores as illustrative; this exists to show the shape and seed a real… See the full description on the dataset page: https://huggingface.co/datasets/davanstrien/glam-extraction-bench-poc.