arthrod/contractscrub-repro

ContractScrub-style synthetic stand-in (unofficial re-implementation)

This is not the ContractScrub dataset and is not affiliated with its authors. The official gold set (tri-fair-lab/contract_scrub) was unreleased as of 2026-10-02. The numbers below come from a scripted synthetic benchmark and cannot be compared with the paper's Table 2.

What this repo contains: a re-implementation of the scoring and per-category prompt structure of ContractScrub: A benchmark for final review of legal contracts (Bang et al., arXiv:2608.20204), run on a stand-in benchmark of 44 CUAD contracts with errors injected by script, and one model run.

Results (synthetic gold; not comparable to the paper)

On this stand-in, Gemma 4 26B (NVFP4) finds 47% of 980 scripted items (macro recall 0.472). The paper reports 0.365 for Gemma 4 26B on its lawyer-written gold. The two figures measure different things.

Precision (0.172) and F1 (0.238) are reported for completeness but are not meaningful here: the CUAD contracts contain pre-existing errors that the paper's lawyers annotated and this gold does not, so a correct flag of a real error counts as a false positive.

Category	Gold items	Recall (exact)	Recall (term-only)	Gold determinable?
defined_terms	555	0.369	0.995	yes
undefined_capitalized_terms	88	0.909	0.977	yes
uncapitalized_defined_terms	39	0.179	0.513	yes
incorrectly_capitalized_terms_in_context	44	0.727	0.773	disputed
unused_defined_terms	88	0.909	0.977	yes
terms_defined_multiple_times	44	0.159	0.795	partly (23 of 44 items name one location twice)
incorrect_section_article_paragraph_references	41	0.000	0.000	no
incorrect_party_references	37	0.081	0.108	no
inconsistent_terms	44	0.909	1.000	yes

Term-only macro recall over the paper's eight term-only categories (its Table 5 has no Inconsistent Language row) is 0.642.[^1] Defined Terms rises from 0.369 exact to 0.995 term-only mostly because the gold records "body" where the prompt tells the model to answer "P" or "Recitals" (see below).

[^1]: Averaging all nine categories, as results.json does, gives 0.682.

Of the 396 outputs, 3 do not parse as JSON (python scrub/score.py); the n_unparsed_outputs: 256 in results.json counted every output that did not begin with {, including fenced JSON that parses.

Known defects of the synthetic gold

• Section references: the "correct" section is chosen at random from the contract, so the right answer cannot be determined from the text; the 0.000 recall is an artifact, not a measurement.
• Party references: the injected clause swaps in the other party with no contextual basis that makes the swap detectable.
• Capitalised in context: the injected clause uses the term generically in lower case and mentions the capitalised form only to say it remains governed by the agreement; it contains no such error.
• Defined multiple times: when the original definition is not found, both gold locations are the injected clause; 23 of the 44 items name one location twice and cannot be matched under exact scoring.
• Locations: gold uses "body" for anything before the first numbered heading and bare numbers inside schedules, contradicting the prompt's location rules.
• Placement: every injected clause is appended after the original text, after the signature block and schedules, as a new numbered section; each contract receives the same fixed sentences.
• Contracts are 44 CUAD agreements chosen by seed, not the paper's 44.
• Model and decoding: nvidia/Gemma-4-26B-A4B-NVFP4 (FP4 quantisation) on vLLM, temperature 0.0; the paper used Google's endpoint and temperature 0.6.
• Prompts are paraphrased from the paper's Appendix G, not verbatim: Undefined Capitalized Terms has 8 exclusions against the paper's 7 and drops "References to agreement sections, exhibits, and schedules"; Terms Defined Multiple Times drops "A term that is merely repeated with the same meaning is NOT an error" and adds a sentence not in the paper; Incorrect References drops the "list each reference as a SEPARATE item" rule and changes the example; the location rules drop the "Attachment X Section Y" rule; an AGREEMENT: trailer is added.

Reproduce

# 1. Re-score the published outputs, no GPU (prints the table above and the diagnostics)
python scrub/score.py --benchmark benchmark/benchmark.jsonl --raw results/raw_outputs.jsonl

# 2. Regenerate the benchmark from CUAD v1 (theatticusproject/cuad, full_contract_txt/)
python scrub/make_synthetic.py --help

# 3. Re-run the model (one A100 80 GB; the published run took 11 minutes, 288 s of generation)
VLLM_USE_FLASHINFER_SAMPLER=0 python scrub/run_eval.py --benchmark benchmark/benchmark.jsonl \
    --out-dir results --model nvidia/Gemma-4-26B-A4B-NVFP4

Versions: requirements.txt (vllm 0.29.0, trackio 0.37.1, huggingface_hub 1.31.0). The published run used Hugging Face Jobs, flavor a100-large, image ghcr.io/astral-sh/uv:python3.12-bookworm, running scrub/job_bootstrap.py with HF_TOKEN passed as a job secret. kv_cache_dtype is forced to bfloat16 because the NVFP4 checkpoint requests an FP8 KV cache the Triton backend did not support on the first (A10G) attempt.

Contents

• SCORING.md: walkthrough of the deterministic multiset scoring
• scrub/prompts.py: the nine categories and prompt assembly (paraphrased from the paper's Appendix G)
• scrub/eval_lib.py: parsing, normalisation, location canonicalisation, multiset scoring
• scrub/make_synthetic.py: the stand-in benchmark generator
• scrub/run_eval.py, scrub/job_bootstrap.py: the model run
• scrub/score.py: re-scoring without a GPU
• tests/: checks that score.py re-derives every number in results.json and that the docs keep their corrections
• benchmark/benchmark.jsonl, results/: the benchmark, raw outputs and results

Attribution and licences

• Contract text: CUAD v1, The Atticus Project, CC BY 4.0, modified here by scripted error injection.
• Category definitions and prompt text: adapted from Bang et al., arXiv:2608.20204, CC BY-NC-SA 4.0.
• Licence for this repo's code and data: not yet set.

@misc{contractscrub2026, title={ContractScrub: A benchmark for final review of legal contracts},
  author={Bang and others}, year={2026}, eprint={2608.20204}, archivePrefix={arXiv}}
@article{hendrycks2021cuad, title={CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review},
  author={Hendrycks, Dan and Burns, Collin and Chen, Anya and Ball, Spencer}, journal={NeurIPS}, year={2021}}