Contra v2 is an educational 41-label English contract-clause classifier fine-tuned from nlpaueb/bert-base-uncased-contracts. The repository ID is rjaloudi/contra-contracts-bert-cuad-clause-classifier.

It is designed as the encoder-routing layer in AI-Enabled Contracts Lab, where classifier candidates remain separate from generative legal analysis. Its release follows modern practices—safe serialization, explicit lineage, grouped-split metrics, per-label limitations, reproducible configuration, and a clean inference path—but this is not a claim of state-of-the-art benchmark performance.

Intended use

Use the model to rank candidate CUAD clause categories for already segmented English contract passages, support error-analysis education, and seed a human-reviewed routing workflow. Preserve top-k labels and scores. Treat every prediction as a hypothesis.

Do not use it as a whole-contract reasoner, an authority source, a substitute for counsel, or an automated decision-maker for enforceability, legal rights, risk, fairness, or required language.

Results

Held-out metric	Value
Accuracy	0.7639
Macro precision	0.6456
Macro recall	0.6151
Macro-F1	0.6036
Test examples	1,008
Labels	41
Zero-F1 labels	6

Training used clipped inverse-square-root class weights with cross-entropy. Checkpoint selection used validation macro-F1; the best validation value was 0.5869. The split is grouped by source contract, so no source contract crosses training, validation, and test partitions. Results describe this project's deterministic classifier split, not the official CUAD extractive-QA benchmark.

Macro-F1 weights each label equally and therefore exposes rare-label failures that accuracy can obscure. Consult evaluation_summary.json for the full metric and per-label disclosure.

Training data

Training examples are deterministic classifier derivatives of CUAD v1, curated by The Atticus Project and declared under CC BY 4.0. Positive annotated answer spans become classifier text and CUAD categories become labels. Exact text/label duplicates are removed before training.

CUAD's publisher states that it makes no representations or warranties about the license status of underlying SEC EDGAR contracts. Preserve attribution and review current CUAD terms before reuse.

How the project uses this model

text

contract passage
  -> Contra v2 top-k clause hypotheses
  -> confidence and abstention policy
  -> independent issue spotting
  -> jurisdiction and prompt/skill routing
  -> local generative LLM
  -> structured result and human review

The project currently has no retrieval-augmented generation layer. The classifier does not supply cases, statutes, regulations, quotations, or verified legal authority.

Limitations and responsible use

This is an educational alpha model, not legal advice.
It classifies individual passages and cannot determine whether a complete contract is adequate.
Six rare labels have zero held-out F1; additional labels remain weak.
Softmax confidence is not calibrated and must not be treated as legal certainty.
The 41-label CUAD taxonomy is closed and incomplete.
English commercial contracts from CUAD do not represent every jurisdiction, document type, drafting style, industry, party, or time period.
Training data may contain annotation artifacts and public-filing language unlike private contracts.
Human review and jurisdiction-specific verification are required for real-world use.

Example

python

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="rjaloudi/contra-contracts-bert-cuad-clause-classifier",
    top_k=3,
)

results = classifier(
    "This Agreement is governed by the laws of the State of New York."
)
print(results)

For private repositories, authenticate with hf auth login before loading.

Reproducibility files

config.json: architecture and exact label mappings.
model.safetensors: model weights using safe serialization.
tokenizer files: preprocessing vocabulary and settings.
training_config.json: project training controls and provenance.
evaluation_summary.json: aggregate and per-label evaluation disclosure.

Licenses

Fine-tuned model bundle: Creative Commons Attribution-ShareAlike 4.0, following the upstream base-model license.
Base model: see the upstream Contracts-BERT model card.
Training data and included dataset derivatives: CC BY 4.0.
AI-Enabled Contracts Lab source code: MIT.

Citation

If you use the training data, cite CUAD:

bibtex

@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}
}