OKTest (FalseReject Re-run): leaderboard
The model answers OKTest's benign questions that contain sensitive trigger words, and the score is the share it refuses, as run by the FalseReject authors. It tests word-level over-sensitivity, where FalseReject and OR-Bench test topic-level look-alikes.
Metric: Rejection rate on benign prompts (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.3) | 11.67 | #6653 |
| 2 | Qwen 2.5 7B Instruct | 12.67 | #2564 |
| 3 | Llama 3.1 405B Instruct | 14.67 | #1115 |
| 4 | Llama 3.3 70B Instruct | 15.67 | #1357 |
| 5 | Llama 3.1 8B Instruct | 21 | #4246 |
| 6 | Claude 3.5 Sonnet (20241022) | 27 | #637 |
| 7 | Claude 3.5 Haiku (20241022) | 38.33 | #1578 |
Interactive version: theaggregate.ai/benchmark?slug=oktest-falsereject-re-run · How It Works · Data refreshed daily, snapshot 2026-11-13.