README.md

---
license: cc-by-4.0
task_categories:
- text-classification
language:
- en
size_categories:
- 10K<n<100K
tags:
- legal
- contracts
- clause-classification
- CUAD
---

# CUAD Clause Classification — with negative ("None Of The Above") samples

This dataset is [`dvgodoy/CUAD_v1_Contract_Understanding_clause_classification`](https://huggingface.co/datasets/dvgodoy/CUAD_v1_Contract_Understanding_clause_classification)
(13,155 attorney-labeled clauses over **41** clause types from 509 commercial
contracts) **with a 42nd reject class added**: `class_id = 41`,
`label = "None Of The Above"`.

The base dataset contains only positively-tagged clauses, so a tagger trained on
it is forced to assign one of the 41 labels to *any* input. The added negatives
let the model learn to say "this is **none** of the 41 clause types."

The format is identical to the source: **7 columns**
(`file_name`, `clause`, `pages`, `class_id`, `label`, `start_at`, `end_at`),
single `train` split.

## Rows

| | count |
|---|--:|
| Positive clauses (class 0–40) | 13,155 |
| **Negatives (class 41)** | **2,600** |
| **Total** | **15,755** |

## How the negatives were built

Two complementary sources, both mapped to `class_id 41`:

1. **In-domain (~82%)** — real, unannotated spans mined from the same 509 source
   contracts (only ~19% of each contract is annotated; the rest is genuine
   contract prose — definitions, confidentiality, indemnification, notices, etc.
   — that is not one of the 41 tracked types). Text is verbatim from the
   contracts. For these rows `file_name` is the source contract.
2. **Out-of-domain (~18%)** — synthetic non-contract sentences (e.g. everyday
   facts, reviews, messages) so the model also rejects clearly non-contract
   input. For these rows `file_name` is empty.

Negative rows use `start_at = end_at = -1` (they are not a single contiguous
annotated span). A light signature filter drops gap chunks that strongly match
one of the 41 clause types, to avoid mislabeling missed annotations as negatives.

## Recommended use

- Train with **per-class F1 / macro-F1**, not raw accuracy (the data is heavily
  imbalanced, ~98:1, and class 41 is now one of the largest classes).
- Optionally add a serving-time confidence threshold (reject when the top
  softmax probability is below τ) as an extra layer over the learned reject class.

## License & attribution

`cc-by-4.0`, inheriting from CUAD v1 (The Atticus Project) and the source dataset.
Please cite CUAD: Hendrycks et al., *CUAD: An Expert-Annotated NLP Dataset for
Legal Contract Review*, NeurIPS 2021 (arXiv:2103.06268).