LarrisX/packed-contracts

Packed contracts

A long context test for decision models: pack several real contracts into one review file and ask about every contract in it, not one needle.

121 commercial contracts from CUAD, each with the questions of Kev's long-document contract benchmark: two "does the contract contain this clause?" questions, a clause-category choice and, where the governing law is a common one, a jurisdiction choice. Labels are CUAD's lawyer annotations. The same contracts are asked alone, then packed into review files of 16k, 32k and 64k tokens, holding about 3, 7 and 14 contracts each. Because every contract is also asked alone, accuracy at each length pairs with the same model's score on the contract by itself.

Packing hurts far less here than on packed-complaints, where a 64k input holds about 246 short complaints: here it holds about 14 long, distinctively named contracts.

Reference scores

Accuracy over all 1,773 questions, scored with Kev's own scorer. Runs are from the code repo; none of them trained on CUAD.

Model	Alone	16k (3)	32k (7)	64k (14)
Kev-0.8B 1.0	0.709	0.683	0.639	0.620
Self-taught on complaints (fresh-32k, mean of 2 runs)	0.712	0.732	0.694	0.667
Self-taught, plus ContractNLI NDAs (fresh-32k-ndas)	0.729	0.735	0.708	0.701
Same, with Kev-27B as the teacher (fresh-32k-ndas-kev27b)	0.744	0.733	0.710	0.732
Trained on real complaint labels (labels-32k)	0.714	0.718	0.694	0.691

Contents

One suite, packed-contracts-v0/, in Kev's format (manifest.json + development.jsonl), scored with kev.benchmark --suite. index.json maps each record to its length and contracts.

Length	Review files	Contracts per file
alone	121	1
16k	37	about 3
32k	17	about 7
64k	8	about 14

207 records in all: a long file's questions are split into groups that share the file, so Kev's scorer fits in GPU memory; every question is still answered on its own. A record is one review file (state) and its questions, each naming its contract by title:

{"state": "Contract review file: 1 commercial contracts filed with the SEC, in no particular order.\n\n===== Contract 1 of 1: 2ThemartComInc_19990826_..._Co-Branding Agreement_ Agency Agreement =====\n...",
 "questions": {"...": {"type": "noul",
                       "instructions": "Consider only the contract titled \"2ThemartComInc_..._Co-Branding Agreement_ Agency Agreement\" in this file. Does one party ...",
                       "label": true}, "...": "..."}}

How it was built

• Contracts. CUAD v1 (theatticusproject/cuad @ a3c393f), at most 7,500 tokens each. Only the contracts that Kev's long-document benchmark (longdoc-v1) assigns to its development partition are used, dealt with Kev's own function, so its locked test contracts are never included. Contracts from one company never share a file.
• Questions and labels. Kev's longdoc-v1 builder (scripts/build_longdoc_v1.py @ a0255ce), with CUAD's lawyer annotations as labels. No model is involved in labelling.
• Packing. Whole contracts packed first-fit into review files filling 70–93% of their nominal length under the Qwen3.5 tokenizer, in a fresh random order at every length.

Licence and credit

CC BY 4.0, as CUAD. Changes from CUAD: contracts selected as above, set in Kev's review-file layout, packed into files of 16k–64k tokens, and given Kev's questions.

• CUAD: Dan Hendrycks, Collin Burns, Anya Chen and Spencer Ball, "CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review", NeurIPS 2021 Datasets and Benchmarks; The Atticus Project. CC BY 4.0.
• Kev by Jared Palmer (Apache-2.0): the review-file layout, the question wording and the partition dealing.