Multi-task Language Understanding on MMLU 0-shot (test)
70.21Humanities Accuracy (0-shot)ZS fine-tuning on V (Value-Matrix Fine-Tuning)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ZS fine-tuning on V (Value-Matrix Fine-Tuning)Fine-tuning objective=Zero-shot (ZS), LoRA adapted parameters=V2026.02 | 70.21 | 76.4 | 65.22 | 66.13 | |
| ZS+FS fine-tuning on V (Value-Matrix Fine-Tuning)Fine-tuning objective=Zero-shot + Few-shot (ZS+FS), LoRA adapted parameters=V2026.02 | 69.62 | 77.96 | 52.25 | 67.48 | |
| ZS fine-tuning on Q/K/VFine-tuning objective=Zero-shot (ZS), LoRA adapted parameters=Q/K/V2026.02 | 69.3 | 75.15 | 68.27 | 63.65 | |
| ZS fine-tuning on KFine-tuning objective=Zero-shot (ZS), LoRA adapted parameters=K2026.02 | 66.42 | 73.99 | 70.76 | 67.14 | |
| ZS fine-tuning on QFine-tuning objective=Zero-shot (ZS), LoRA adapted parameters=Q2026.02 | 66.05 | 71.77 | 67.42 | 60.83 | |
| Base model (Qwen2.5-3B-Instruct)Fine-tuning=None2026.02 | 63.62 | 70.12 | 70.69 | 63.52 |