CLAMBER - Few-shot - Identification F1: leaderboard
CLAMBER (ACL 2024) evaluates detecting ambiguous information needs and generating clarifying questions. Few-shot - Identification F1 under the named prompting protocol. Metric: F1 (%). Helpfulness is human-rated; BERTScore measures similarity to reference questions and is not task success. Source: arXiv 2405.12063v2, tables 3 and 7. Chat and instruct models remain distinct; ChatGPT here is GPT-3.5-Turbo-16k.
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 70B Chat | 39.64 | #1275 |
| 2 | vicuna-13B | 36.62 | #1450 |
| 3 | Llama 2 13B Chat | 35.08 | #1414 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=clamber-few-shot-identification-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.