AbstentionBench - subjective - KUQ/Controversial - Precision: leaderboard
Metric: Abstention Precision (%). Source: github.com. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek R1 Distill Llama 70B | 96.6 |
| 2 | Llama-3.1-Tulu-3-8B-SFT | 96.15 |
| 3 | Llama-3.1-Tulu-3-8B | 94.52 |
| 4 | Llama 3.3 70B Instruct | 93.09 |
| 5 | Llama-3.1-Tulu-3-8B-DPO | 92.41 |
| 6 | O1 | 92.21 |
| 7 | Mistral 7B Instruct (v0.3) | 91.58 |
| 8 | GPT-4o | 90.8 |
| 9 | Llama-3.1-Tulu-3-70B-DPO | 90.6 |
| 10 | Llama 3.1 70B Instruct | 90.37 |
| 11 | Llama-3.1-Tulu-3-70B | 89.96 |
| 12 | Llama 3.1 70B | 88.89 |
| 13 | Llama 3.1 405B Instruct | 88.52 |
| 14 | Gemini 1.5 Pro | 86.69 |
| 15 | Llama 3.1 8B | 86.41 |
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-subjective-kuq-controversial-precision · How It Works · Data refreshed daily, snapshot 2026-09-19.