MedErrBench (Arabic) - Error Correction: leaderboard
Metric: Correction ROUGE-1 (%): unigram overlap of the corrected sentence with the clinician reference (ROUGE-chinese tokenizer for Chinese); 97 Arabic test notes adapted from MedArabiQ and MedAraBench questions (53 with an error, 44 without) with clinician-verified inserted errors; main-table prompting (the paper does not state its shot setting); printed as fractions and shown times 100; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Doubao-1.5-Thinking-Pro | 58.2 | #133 |
| 2 | MedGemma-4B | 43.8 | #731 |
| 3 | Llama 3.3 70B Instruct | 41.2 | #520 |
| 4 | GPT-4o | 39.9 | #333 |
| 5 | Qwen 2.5 7B Instruct | 32.9 | #846 |
| 6 | Gemini 2.0 Flash | 31.5 | #331 |
| 7 | Llama 3.1 8B Instruct | 31.1 | #1018 |
| 8 | Gemini 2.5 Flash Lite | 30.3 | #413 |
| 9 | GPT-4o Mini | 26 | #588 |
| 10 | ALLaM-7B-Instruct-preview | 4.5 | #698 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mederrbench-arabic-error-correction · How It Works · Data refreshed daily, snapshot 2026-10-11.