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.

Related repos

Repo	Role
arthrod/contractscrub-repro-trackio	Space: Trackio dashboard for the published model run
theatticusproject/cuad	Source of the 44 contract texts (CUAD v1)
arthrod/sec-ex10-exhibits	Sibling legal corpus: SEC EX-10 material contracts as Markdown (CUAD's contracts are also EX-10 exhibits)
arthrod/neurotic_docx_bench	Sibling legal benchmark: DOCX redline/conversion bench over contracts
arthrod/docx-corpus	Sibling corpus: classified web .docx documents

Configs

All configs have a single test split.

Config	Rows	What one row is
benchmark (default)	44	One contract: full text + gold issues for all nine categories (same content as benchmark/benchmark.jsonl)
gold_items	980	One gold issue, flattened (contract, category, fields) — easiest to filter in the viewer
model_outputs	396	One (contract, category) model response from the published run (44 × 9; same as results/raw_outputs.jsonl)
results	20	Scores per category and macro, for exact and term_only scoring (same as results/results.json)

The JSONL/JSON originals are kept because scrub/score.py and tests/ read them; the parquet files are built from them by a deterministic conversion and carry the same content.

Schema

benchmark

Column	Type	Description
contract_id	string	CUAD file stem (filer, filing date, form, exhibit number, agreement type)
text	string	Full contract text with the injected clauses appended (18,663–60,445 chars, median 37,857)
n_chars	int32	len(text)
gold	struct	One list per category. defined_terms, undefined_capitalized_terms, uncapitalized_defined_terms, incorrectly_capitalized_terms_in_context, unused_defined_terms: list<{term, location}>; terms_defined_multiple_times: list<{term, location1, location2}>; incorrect_section_article_paragraph_references, incorrect_party_references: list<{location, wrong, correct}>; inconsistent_terms: list<{section1, section2}>

gold_items

Column	Type	Description
contract_id	string	Joins to benchmark.contract_id
category	string	One of the nine categories
item_idx	int32	Position within that contract's category list
term, location, location1, location2, wrong, correct, section1, section2	string (nullable)	The item's fields; only the fields used by that category are set

Gold items per category: defined_terms 555, undefined_capitalized_terms 88, unused_defined_terms 88, incorrectly_capitalized_terms_in_context 44, terms_defined_multiple_times 44, inconsistent_terms 44, incorrect_section_article_paragraph_references 41, uncapitalized_defined_terms 39, incorrect_party_references 37.

model_outputs

Column	Type	Description
contract_id	string	Joins to benchmark.contract_id
category	string	The category prompt that produced this output
output	string	Raw model text (sometimes fenced in a ```json block)
n_items_parsed	int32	Items recovered by scrub/eval_lib.parse_json_objects (0 if unparseable)

results

Column	Type	Description
model	string	nvidia/Gemma-4-26B-A4B-NVFP4
scoring	string	exact (all fields) or term_only (Appendix C ablation)
category	string	A category, or MACRO (9 categories)
precision, recall, f1	float64	Pooled over the 44 contracts
tp, fp, fn, n_gold	int32	Counts (null on the macro rows)

Examples

benchmark (text truncated, first two items per category):

{
  "contract_id": "RaeSystemsInc_20001114_10-Q_EX-10.57_2631790_EX-10.57_Co-Branding Agreement",
  "text": "CO-BRANDING AGREEMENT  (FORM)\n\n This CO-BRANDING AGREEMENT (the \"Agreement\") is made and entered into as of Nov 5, 1999, (the \"Effective Date\") by and between NETTAXI Online Comm…",
  "n_chars": 44596,
  "gold": {
    "defined_terms": [{"term": "Impressions", "location": "1.3"}, {"term": "Intellectual Property Rights", "location": "1.4"}],
    "undefined_capitalized_terms": [{"term": "Acceptance Authority", "location": "11.1"}, {"term": "Vendor Handbook", "location": "11.2"}],
    "uncapitalized_defined_terms": [{"term": "Intellectual Property Rights", "location": "3.4"}],
    "incorrectly_capitalized_terms_in_context": [{"term": "Impressions", "location": "11.3"}],
    "unused_defined_terms": [{"term": "Renewal Option Fee", "location": "11.4"}, {"term": "Disaster Recovery Plan", "location": "11.5"}],
    "terms_defined_multiple_times": [{"term": "Confidential Information", "location1": "6.2", "location2": "11.6"}],
    "incorrect_section_article_paragraph_references": [{"location": "11.7", "wrong": "7.1", "correct": "8.1"}],
    "incorrect_party_references": [{"location": "11.8", "wrong": "Inc., a Delaware corporation with principal offices at 2165 S. Bascom Avenue", "correct": "NETTAXI Online Communities"}],
    "inconsistent_terms": [{"section1": "11.9", "section2": "11.10"}]
  }
}

gold_items:

{"contract_id": "RaeSystemsInc_…_Co-Branding Agreement", "category": "terms_defined_multiple_times", "item_idx": 0, "term": "Confidential Information", "location": null, "location1": "6.2", "location2": "11.6", "wrong": null, "correct": null, "section1": null, "section2": null}
{"contract_id": "RaeSystemsInc_…_Co-Branding Agreement", "category": "inconsistent_terms", "item_idx": 0, "term": null, "location": null, "location1": null, "location2": null, "wrong": null, "correct": null, "section1": "11.9", "section2": "11.10"}

model_outputs and results:

{"contract_id": "RaeSystemsInc_…_Co-Branding Agreement", "category": "undefined_capitalized_terms", "output": "```json\n{\"undefined_capitalized_terms\": [{\"term\": \"Acceptance Authority\", \"location\": \"11.1\"}, {\"term\": \"Vendor Handbook\", \"location\": \"11.2\"}, {\"term\": \"Goods\", \"location\": \"11.8\"}, …", "n_items_parsed": 5}
{"model": "nvidia/Gemma-4-26B-A4B-NVFP4", "scoring": "exact", "category": "defined_terms", "precision": 0.1647, "recall": 0.3694, "f1": 0.2278, "tp": 205, "fp": 1040, "fn": 350, "n_gold": 555}

Usage

from datasets import load_dataset

bench = load_dataset("arthrod/contractscrub-repro", "benchmark", split="test")
gold = load_dataset("arthrod/contractscrub-repro", "gold_items", split="test").to_polars()
print(gold.group_by("category").len())

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.

Limitations: 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. The extracted "other party" string can also be a fragment of the recitals rather than a party name (first contract: wrong = "Inc., a Delaware corporation with principal offices at 2165 S. Bascom Avenue").
• 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 (also as parquet under data/, see Configs)

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}}