AbstentionBench - underspecified intent - SituatedQA/Geo - F1: leaderboard
AbstentionBench evaluates whether a model appropriately declines unanswerable or underspecified questions. This slice reports underspecified intent - SituatedQA/Geo - 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 | GPT-4o | 70.99 | #357 |
| 2 | O1 | 66.44 | #190 |
| 3 | Llama 3.3 70B Instruct | 64.46 | #541 |
| 4 | Llama-3.1-Tulu-3-70B-DPO | 64.04 | #732 |
| 5 | Llama-3.1-Tulu-3-70B | 63.28 | #499 |
| 6 | Llama 3.1 8B Instruct | 63.08 | #1017 |
| 7 | Llama-3.1-Tulu-3-8B | 62.66 | #848 |
| 8 | Llama-3.1-Tulu-3-8B-DPO | 60.53 | #992 |
| 9 | Mistral 7B Instruct (v0.3) | 59.85 | #1255 |
| 10 | Qwen 2.5 32B Instruct | 58.68 | #509 |
| 11 | Llama-3.1-Tulu-3-8B-SFT | 56.72 | #977 |
| 12 | Llama 3.1 405B Instruct | 54.23 | #470 |
| 13 | Llama 3.1 70B Instruct | 54.18 | #567 |
| 14 | DeepSeek R1 Distill Llama 70B | 53.48 | #614 |
| 15 | OLMo-7B-0724-Instruct-hf | 53.11 | #1441 |
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-intent-situatedqa-geo-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.