Qwen2.5-Math-7B-CFT: benchmark results
TIGER-Lab's Qwen2.5-Math-7B-CFT, built on Qwen2.5-Math-7B and trained with Critique Fine-Tuning on the 50,000-sample WebInstruct-CFT dataset, teaching the model to critique responses rather than imitate them. Provider: TIGER-Lab. Released 2025-01-30. Access: Open.
Aggregate Rating 1321 ± 25, rank #1438 of 1656 rated models, from 184 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard - MATH Level 5 | 55.74 | Score | 99.3 |
| Open Arabic LLM - Arabic MMLU Civics (High School) | 47.13 | Accuracy (%) | 73.8 |
| Open Japanese LLM - MR | 82.2 | Score (%) | 49.5 |
| Open Portuguese LLM - ASSIN2 RTE | 87.82 | Macro F1 (%) | 42.3 |
| Open Arabic LLM - Madinah QA Arabic Language (Grammar) | 30.68 | Accuracy (%) | 41 |
| Open LLM Leaderboard - GPQA | 28.61 | Score | 41 |
| Open LLM Leaderboard - BBH | 46.37 | Score | 37.2 |
| Open Japanese LLM - Wiki Coreference SET F1 | 1.61 | Score (%) | 35.7 |
| Open LLM Leaderboard - MMLU-Pro | 29.45 | Score | 34.6 |
| Open LLM Leaderboard - MuSR | 38.87 | Score | 34.6 |
| Open Portuguese LLM - OAB Exams | 35.81 | Accuracy (%) | 33.1 |
| Open Portuguese LLM - BLUEX | 40.75 | Accuracy (%) | 33 |
Interactive version: theaggregate.ai/model?slug=qwen2-5-math-7b-cft · How It Works · Data refreshed daily, snapshot 2026-10-20.