magnum-v2.5-12B-kto: benchmark results
anthracite-org's magnum-v2.5-12B-kto, a KTO+DPOP fine-tune of their magnum-12B-v2 aimed at replicating Claude 3 Sonnet and Opus prose style. Provider: anthracite-org. Released 2024-08-12. Access: Open.
Aggregate Rating 1470 ± 18, rank #874 of 1635 rated models, from 70 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - Ko-EQ-Bench | 45.7 | EQ-Bench score | 94.1 |
| Open Korean LLM - Ko-Winogrande | 67.88 | Accuracy (%) | 92 |
| Open Korean LLM Leaderboard | 43.8 | Average Score (%) | 89.1 |
| Open Japanese LLM - Wiki Dependency SET F1 | 39.07 | Score (%) | 82 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 47.84 | Exact match, flexible extract (%) | 81.7 |
| Open Japanese LLM - Mbpp Pylint Check | 88.35 | Score (%) | 74.5 |
| Open Korean LLM - Ko-IFEval | 35.5 | Strict accuracy, prompt/instruction mean (%) | 73.1 |
| Open Japanese LLM - Xlsum JA RougeLsum | 24.38 | Score (%) | 71.4 |
| Open Japanese LLM - Xlsum JA Rouge1 | 28.25 | Score (%) | 69.7 |
| Open Korean LLM - Ko-GPQA-Diamond | 24.75 | Normalized accuracy (%) | 69.5 |
| Open Japanese LLM - CG | 46.59 | Score (%) | 68.6 |
| Open Korean LLM - KorNAT-Harmless | 63.85 | Normalized accuracy (%) | 67.5 |
Interactive version: theaggregate.ai/model?slug=magnum-v2-5-12b-kto · How It Works · Data refreshed daily, snapshot 2026-11-05.