AbstentionBench - false premise - QAQA - Precision: leaderboard
Metric: Abstention Precision (%). Source: github.com. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek R1 Distill Llama 70B | 92.86 |
| 2 | GPT-4o | 92.22 |
| 3 | O1 | 88.35 |
| 4 | Llama 3.1 70B Instruct | 85.71 |
| 5 | Llama 3.1 405B Instruct | 84.62 |
| 6 | Gemini 1.5 Pro | 83.85 |
| 7 | Llama 3.3 70B Instruct | 83.69 |
| 8 | Qwen 2.5 32B Instruct | 83.02 |
| 9 | Llama-3.1-Tulu-3-70B | 82.72 |
| 10 | Llama-3.1-Tulu-3-70B-DPO | 81.87 |
| 11 | Mistral 7B Instruct (v0.3) | 81.41 |
| 12 | Llama 3.1 8B | 81 |
| 13 | Llama-3.1-Tulu-3-8B | 77.07 |
| 14 | Llama-3.1-Tulu-3-8B-DPO | 76.97 |
| 15 | Llama 3.1 70B | 75.23 |
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-false-premise-qaqa-precision · How It Works · Data refreshed daily, snapshot 2026-09-19.