246 questions on 20 redlined contract PDFs. They evaluate document parsers, not models: can a parser give a downstream agent what it needs to answer a question about a heavily negotiated contract?

Rows carry the answer key, so they are not contamination-free. The scored board is a separate set that is not distributed.

Columns

The same columns as OB-LegalQA, so a harness written against the public sample runs on this unchanged.

column	meaning
`id`	stable item id
`question`	the question put to the agent
`answer`	gold answer, from the position the parties agreed
`answer_detail`	why that is the answer
`task_type`	which redline pattern the question turns on
`documents`	contract id(s) the question is about
`pdf`	path(s) to the digital PDF in this repo
`pdf_scanned`	path(s) to the same contract as page images, no text layer
`match_correct`	strings that indicate the agent answered correctly
`match_stale`	strings that indicate it answered from struck text
`match_fused`	strings that indicate it merged struck and surviving text

Unlike the public sample, this split does not ship views/. Those renderings are the ideal parse of each contract, and the contracts here overlap the scored set.

Provenance

Contracts come from crosbylegal/RedlineBench (CC-BY-4.0), converted to PDF. Derived from RedlineBench by Crosby Legal, adapted. Licensed CC-BY-4.0.

Benchmark design: openbenchmarks.com/doc-processing.

Downloads last month: 36

Total file size:

357 MB