README.md

metadata

license: apache-2.0
configs:
  - config_name: extract-bench
    features:
      - name: id
        dtype: string
      - name: category
        dtype: string
      - name: pdf
        dtype: string
      - name: data_schema
        dtype: string
      - name: expected_output
        dtype: string
      - name: field_rules
        dtype: string
      - name: repeated_structure
        dtype: string
      - name: tags
        sequence: string
    data_files:
      - split: short
        path: short.jsonl
      - split: medium
        path: medium.jsonl
      - split: long
        path: long.jsonl
language:
  - en
pretty_name: ExtractBench
size_categories:
  - n<1K
tags:
  - document-extraction
  - structured-extraction
  - information-extraction
  - pdf
  - benchmark
  - evaluation
  - json-schema
  - visual-grounding
  - forms
  - tables
citation: |
  @misc{extractbench2026,
        title={ExtractBench: A Benchmark for Schema-Guided Enterprise Document Extraction},
        author={LlamaIndex},
        year={2026},
        url={https://github.com/run-llama/ExtractBench},
  }

ExtractBench — smoke subset

This is the test-data branch: a 6-document sample (3 short, 2 medium, 1 long) for quick smoke runs. The benchmark itself lives on the main branch — 370 documents across 4,869 pages — and that is what any reported score must be run against.

The rows here are byte-identical to their counterparts on main, in the same format and with the same tags; only the number of documents differs. Use it to check that a pipeline is wired up correctly before spending a full run.

uv run extract-bench download --test          # -> data/test/<split>/
uv run extract-bench run <pipeline> --test    # inference -> evaluation -> reports

Split	File	Documents	Example
Short	short.jsonl	3	scanned tax return, handwritten-annotated disposal permit, pivoted election results
Medium	medium.jsonl	2	district check register, quarterly earnings deck
Long	long.jsonl	1	66-page clinical measurement listing

The sample spans eight challenge tags (long lists, cross-page continuation, pivoted tables, packed cells, needle-in-haystack, dense forms, filer-versus-reviewer edits), scanned and handwritten capture, and four business domains.

Five of the six documents carry word-level bounding boxes in their ground truth, so every split reports grounding as well as value accuracy. The earnings deck has value-level ground truth only; grounding metrics are omitted for it, as they are for any document whose ground truth carries no boxes.

• Code: run-llama/ExtractBench
• Full dataset: main branch