| # Data |
|
|
| ## CUAD (Contract Understanding Atticus Dataset) |
|
|
| **Contains:** |
|
|
| - 510 commercial contracts (NDAs, service agreements, licensing agreements, etc.) |
| - 13,000+ expert annotations |
| - 41 clause type categories (e.g., "Governing Law", "Termination for Convenience", "Limitation of Liability", "Non-Compete", "Indemnification", etc) |
| - Each annotation is a span within a contract + the question it answers |
|
|
| **Uses:** |
|
|
| 1. **For Classification Agent taxonomy**: the 41 clause categories are the classification schema. When the classification agent receives a clause segment, it uses the 41 types as the set of all labels it can assign. Approach: put all 41 category names + brief descriptions in the system prompt as a reference list |
|
|
| 2. **For Risk Analysis Agent**: The annotations tell which categories expert annotators were trained to flag as legally significant, which can support the risk scoring logic (gives LLM structured guidance about what to look for) |
|
|
| 3. **Benchmark corpus (secondary)**: the 510 CUAD contracts can be embedded and added to the vector store alongside EDGAR contracts |
|
|
| 4. Sample input contracts for testing/demo (maybe): Since CUAD contracts are real commercial contracts, a subset (20-30?) could be used as "uploaded contracts" during testing and demo, rather than needing to find separate test documents. |
|
|
|
|
| **Need to collect:** |
|
|
| - Full taxonomy: list of 41 category names + description of each -> data/cuad/taxonomy.json |
| - Contract texts: save the 510 contract full texts -> data/cuad/contracts/ |
| - Annotations (the labeled spans per contract) -> data/cuad/annotations.json |
