Commonsense Reasoning on LLM-adapters Reasoning (BoolQ, PIQA, SIQA, HellaSwag, ARC-c)
76.19AccuracyLoRA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LoRABackbone=Llama-3.2-1B, Trainable Parameters (%)=0.452026.04 | 76.19 | 19.06 | |
| LARS-LPBackbone=Llama-3.2-1B, Trainable Parameters (%)=0.672026.04 | 75.97 | 16.07 | |
| AdaLoRABackbone=Llama-3.2-1B, Trainable Parameters (%)=0.452026.04 | 75.8 | 16.06 | |
| LARSBackbone=Llama-3.2-1B, Trainable Parameters (%)=0.672026.04 | 75.6 | 13.55 | |
| IA3Backbone=Llama-3.2-1B, Trainable Parameters (%)=0.012026.04 | 71.66 | 19.83 | |
| PromptBackbone=Llama-3.2-1B, Trainable Parameters (%)=0.0032026.04 | 40.79 | 14.34 | |
| LARS-LPBackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.672026.04 | 39.32 | 32.39 | |
| PrefixBackbone=Llama-3.2-1B, Trainable Parameters (%)=0.032026.04 | 39.23 | 13.47 | |
| AdaLoRABackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.452026.04 | 38.4 | 32.06 | |
| IA3Backbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.012026.04 | 37.44 | 36.99 | |
| LARSBackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.672026.04 | 37.33 | 27.44 | |
| LoRABackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.452026.04 | 35.73 | 35.72 | |
| PromptBackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.0032026.04 | 15.47 | 29.31 | |
| PrefixBackbone=Qwen2.5-7B-Instruct, Trainable Parameters (%)=0.032026.04 | 9.03 | 25.49 |