General Language Understanding on MMLU (Accuracy)
72.45MMLU AccuracyQwen3-4B
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3-4BVersion=25072026.04 | 72.45 | |
| UnalignedModel Family=Qwen3-8B2026.05 | 70.8 | |
| Curri-DPOModel Family=Qwen3-8B2026.05 | 70 | |
| SequentialModel Family=Qwen3-8B2026.05 | 69.8 | |
| Standard DPOModel Family=Qwen3-8B2026.05 | 69.7 | |
| Sqrt-CompetenceModel Family=Qwen3-8B2026.05 | 69.6 | |
| SpB2.0-5BSFT Strategy=General SFT2026.04 | 69.06 | |
| Staged-CompetenceModel Family=Qwen3-8B2026.05 | 68.2 | |
| Qwen2.5-3B2026.04 | 66.43 | |
| SpB1.0-7B2026.04 | 66.34 | |
| UnalignedModel Family=Yi-1.5-9B2026.05 | 66 | |
| Standard DPOModel Family=Yi-1.5-9B2026.05 | 63.5 | |
| SequentialModel Family=Yi-1.5-9B2026.05 | 62.2 | |
| Curri-DPOModel Family=Yi-1.5-9B2026.05 | 61.8 | |
| General Safe-RLHFTraining Configuration=General Safe-RLHF2026.04 | 61 | |
| Sqrt-CompetenceModel Family=Yi-1.5-9B2026.05 | 60.3 | |
| Llama3.2-3B2026.04 | 60.08 | |
| Staged-CompetenceModel Family=Yi-1.5-9B2026.05 | 59.4 | |
| Gemma-3-4B2026.04 | 58.44 | |
| Original ModelTraining Configuration=Original Model2026.04 | 57 | |
| ARES (Ministral)Training Configuration=Ours (Ministral)2026.04 | 57 | |
| ARES (Qwen)Training Configuration=Ours (Qwen)2026.04 | 56 | |
| UnalignedModel Family=LLaMA-3-8B2026.05 | 56 | |
| Sqrt-CompetenceModel Family=LLaMA-3-8B2026.05 | 55.9 | |
| Standard DPOModel Family=LLaMA-3-8B2026.05 | 53.9 | |
| SequentialModel Family=LLaMA-3-8B2026.05 | 53 | |
| Staged-CompetenceModel Family=LLaMA-3-8B2026.05 | 52.6 | |
| Curri-DPOModel Family=LLaMA-3-8B2026.05 | 52.1 | |
| Initial RLHFTraining Configuration=Initial RLHF2026.04 | 48 |