AbstentionBench - underspecified context - GSM8K - F1: leaderboard
AbstentionBench evaluates whether a model appropriately declines unanswerable or underspecified questions. This slice reports underspecified context - GSM8K - 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. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 (High) | 97.68 | #190 (O1) |
| 2 | O1 | 97.6 | #190 |
| 3 | Llama 3.1 405B Instruct | 97.34 | #470 |
| 4 | Qwen 2.5 32B Instruct | 97.16 | #509 |
| 5 | Llama 3.1 70B Instruct | 97.03 | #567 |
| 6 | Gemini 1.5 Pro | 96.99 | #454 |
| 7 | Llama 3.1 8B Instruct | 96.07 | #1017 |
| 8 | GPT-4o | 95.84 | #357 |
| 9 | Llama 3.3 70B Instruct | 95.34 | #541 |
| 10 | Llama-3.1-Tulu-3-70B-DPO | 94.25 | #732 |
| 11 | Llama-3.1-Tulu-3-70B | 93.5 | #499 |
| 12 | Mistral 7B Instruct (v0.3) | 91.55 | #1255 |
| 13 | Llama-3.1-Tulu-3-8B-DPO | 85.86 | #992 |
| 14 | Llama-3.1-Tulu-3-8B | 85.4 | #848 |
| 15 | Llama-3.1-Tulu-3-8B-SFT | 81.27 | #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-underspecified-context-gsm8k-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.