PHTest (FalseReject Re-run): leaderboard
The model answers up to 1,000 PHTest pseudo-harmful prompts, generated to look harmful while being benign, and the score is the share it refuses, as run by the FalseReject authors. PHTest's own paper reports these results only in figures, so this re-run is the only numeric table.
Metric: Rejection rate on benign prompts (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.3 70B Instruct | 5.8 | #1357 |
| 2 | Mistral 7B Instruct (v0.3) | 6.3 | #6653 |
| 3 | Qwen 2.5 7B Instruct | 6.5 | #2564 |
| 4 | Llama 3.1 405B Instruct | 10.2 | #1115 |
| 5 | Llama 3.1 8B Instruct | 11.6 | #4246 |
| 6 | Claude 3.5 Sonnet (20241022) | 25.6 | #637 |
| 7 | Claude 3.5 Haiku (20241022) | 39.8 | #1578 |
Interactive version: theaggregate.ai/benchmark?slug=phtest-falsereject-re-run · How It Works · Data refreshed daily, snapshot 2026-11-13.