Skip to content

Dataset · llamaindex

Document data extraction benchmark

llamaindex/ExtractBench

Documents and schemas for testing structured JSON extraction with evidence.

What's inside

ExtractBench tests systems that take a full document and a user schema to return structured data with evidence. Inputs include a born-digital or scanned document and a schema, and outputs are expected to be a schema-valid JSON with page numbers and bounding boxes for each value. Systems must return full correct values, including repeating entries, use null for missing information, and ground extracted data to the source. Language is English, size is under 1k examples.

What you can do with it

  • Evaluating LLMs and document-AI on JSON extraction
  • Testing schema-guided extraction with evidence
  • Testing OCR pipelines on scans
Languages
English
Size
up to 1K rows
Downloads this month
20.5K
Likes
74

How to load it

from datasets import load_dataset

ds = load_dataset("llamaindex/ExtractBench")