Polish Legal RAG - Passage QA: leaderboard
Polish Legal RAG leaderboard by gaius-lex, sum_rag task: answering Polish legal questions strictly from provided passages, each answer graded 0-2 by a GPT-4o judge and scored as F1, 0-100, higher is better. Near ceiling (83-100); results frozen since August 2025.
Metric: F1 of GPT-4o judge grades (0-100). Source: huggingface.co. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 (Thinking) | 100 | #263 |
| 2 | GPT-5 Mini | 99.45 | #193 |
| 3 | Claude Sonnet 4 | 99.45 | #207 |
| 4 | DeepSeek V3 (0324) | 99.45 | #349 |
| 5 | GLM-4.5 | 99.45 | #283 |
| 6 | GLM 4.5 Air | 99.09 | #403 |
| 7 | Qwen 3 14B | 98.62 | #542 |
| 8 | GPT-5 | 98.53 | #105 |
| 9 | Qwen 3 32B | 98.15 | #449 |
| 10 | Qwen 3 30B A3B | 98.06 | #496 |
| 11 | GPT-OSS-120B | 97.87 | #346 |
| 12 | Llama 4 Maverick Instruct FP8 | 97.68 | #424 |
| 13 | Llama 4 Maverick Instruct | 97.68 | #467 |
| 14 | Llama 4 Scout Instruct | 97.3 | #547 |
| 15 | Kimi K2 | 97.3 | #252 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=polish-legal-rag-passage-qa · How It Works · Data refreshed daily, snapshot 2026-11-05.