AbstentionBench - false premise - QAQA - F1: leaderboard
AbstentionBench evaluates whether a model appropriately declines unanswerable or underspecified questions. This slice reports false premise - QAQA - F1 as Abstention F1 (%); values converted from 0-1 to percent. Precision and recall must be read together: always abstaining is not necessarily useful. Scenario/dataset pairs and SFT/DPO/RL checkpoints remain separate; no synthetic overall average. Source: facebookresearch/AbstentionBench analysis/abstention_performance.csv.
Metric: Abstention F1 (%). Source: github.com. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 74.13 | #190 |
| 2 | GPT-4o | 68.14 | #357 |
| 3 | Gemini 1.5 Pro | 67.51 | #454 |
| 4 | Llama-3.1-Tulu-3-70B | 66.39 | #499 |
| 5 | Llama-3.1-Tulu-3-70B-DPO | 63.81 | #732 |
| 6 | Llama 3.1 405B Instruct | 63 | #470 |
| 7 | Qwen 2.5 32B Instruct | 59.46 | #509 |
| 8 | Llama-3.1-Tulu-3-8B-DPO | 59.18 | #992 |
| 9 | OLMo-7B-0724-Instruct-hf | 59.02 | #1441 |
| 10 | Mistral 7B Instruct (v0.3) | 57.6 | #1255 |
| 11 | Llama 3.3 70B Instruct | 55.4 | #541 |
| 12 | Llama-3.1-Tulu-3-8B | 54.75 | #848 |
| 13 | Llama 3.1 70B Instruct | 54.55 | #567 |
| 14 | Llama 3.1 8B Instruct | 53.39 | #1017 |
| 15 | Llama-3.1-Tulu-3-8B-SFT | 47.8 | #977 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-false-premise-qaqa-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.