Legal Hallucination Detection on CUAD Matched-subset 120 contracts (test)
7.6FARllama-3.3-70b
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| llama-3.3-70brun_id=12026.06 | 7.6 | 17.9 | 89.2 | 63.4 | 36.3 | 3.6 | |
| gemma-debaterun_id=1, Intervention=typed debate pipeline2026.06 | 8.4 | 14.4 | 89.7 | 58.6 | 43.3 | 2.4 | |
| gpt-5.2run_id=12026.06 | 10.7 | 12 | 88.9 | 61 | 43.7 | 2.6 | |
| qwen3-32brun_id=1, Note=Row-count caveat: 4,817 rows vs 4,920 nominal2026.06 | 14.9 | 10.6 | 86.5 | 64.6 | 43.4 | 3.4 | |
| gemma-baserun_id=1, Intervention=none (base model)2026.06 | 15.4 | 15.8 | 84.5 | 64.8 | 41.8 | 5.2 | |
| gemini-3-flashrun_id=12026.06 | 19.1 | 4.5 | 85.4 | 66.2 | 46.8 | 3.8 |