Instruction-following Evaluation Suite (MMLU, DROP, HEval, BBH) (test)
79.67MMLUBi-LoRA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Bi-LoRACost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.67 | 56.49 | 71.34 | 58.93 | 66.61 | |
| LoRA-LookSAMCost=×1.5, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.63 | 56.35 | 67.88 | 58.41 | 65.57 | |
| LoRA-ESAMCost=×1.5, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.55 | 54.83 | 65.04 | 58.43 | 64.46 | |
| Flat-LoRACost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.51 | 54.9 | 71.54 | 58.13 | 66.02 | |
| LoRA-oBARCost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.37 | 55.7 | 69.92 | 58.53 | 65.88 | |
| LoRA-nBARCost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.34 | 55.96 | 69.92 | 58.67 | 65.97 | |
| VanillaCost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.28 | 51.55 | 70.93 | 57.47 | 64.81 | |
| ROPCost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.22 | 52.74 | 71.34 | 57.84 | 65.28 | |
| LoRA-SAMCost=×2, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.19 | 54.54 | 70.53 | 58.25 | 65.63 | |
| RWP_LoRACost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 79.07 | 51.69 | 71.34 | 57.45 | 64.89 | |
| RWP_fullCost=×1, Backbone=Qwen 2.5-14B, Evaluation Protocol=Fine-tuning2025.08 | 78.92 | 54.75 | 71.13 | 57.76 | 65.64 |