AbstentionBench - stale - FreshQA - F1: leaderboard
AbstentionBench evaluates whether a model appropriately declines unanswerable or underspecified questions. This slice reports stale - FreshQA - 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. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 8B | 52.08 | #1130 |
| 2 | Llama-3.1-Tulu-3-8B | 51.57 | #848 |
| 3 | Llama 3.1 70B | 50.63 | #588 |
| 4 | Llama-3.1-Tulu-3-8B-DPO | 48.78 | #992 |
| 5 | Llama-3.1-Tulu-3-70B | 48.55 | #499 |
| 6 | Llama-3.1-Tulu-3-70B-DPO | 48.48 | #732 |
| 7 | GPT-4o | 47.42 | #357 |
| 8 | OLMo-7B-0724-Instruct-hf | 47.19 | #1441 |
| 9 | Llama 3.3 70B Instruct | 45.8 | #541 |
| 10 | Qwen 2.5 32B Instruct | 45.4 | #509 |
| 11 | O1 | 43.04 | #190 |
| 12 | Gemini 1.5 Pro | 42.5 | #454 |
| 13 | Llama 3.1 405B Instruct | 42.24 | #470 |
| 14 | Llama-3.1-Tulu-3-8B-SFT | 42.02 | #977 |
| 15 | Llama 3.1 70B Instruct | 40 | #567 |
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-stale-freshqa-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.