CitizenQuery-UK (Few-Shot): leaderboard
Metric: Mean F1@K (0-1) on the first public split of CitizenQuery-UK (7,355 synthetic citizen questions about UK government services, each with an expected answer generated from gov.uk pages); answers the abstention classifier flags are set aside; Llama-3.1-8B-Instruct splits each answer into atomic claims and checks each against the expected answer's atomic facts, and F1@K is the harmonic mean of precision and recall up to K, the number of expected facts; temperature 0, at most 1,024 output tokens; three example query-response pairs from earlier dataset iterations in the prompt; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 0.86 | #271 |
| 2 | Qwen 3 32B | 0.86 | #424 |
| 3 | Llama 3.3 70B Instruct | 0.84 | #520 |
| 4 | GPT-5.2 | 0.82 | #105 |
| 5 | Gemini 3 Flash | 0.8 | #93 |
| 6 | GPT-4.1 | 0.8 | #240 |
| 7 | Kimi K2 | 0.8 | #236 |
| 8 | Llama 3.1 8B Instruct | 0.8 | #1018 |
| 9 | GPT-4o | 0.8 | #333 |
| 10 | O3 | 0.78 | #121 |
| 11 | GPT-OSS-20B | 0.75 | #499 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=citizenquery-uk-few-shot · How It Works · Data refreshed daily, snapshot 2026-10-11.