Model Efficiency Benchmarking on Llama 3.1 8B
62.4Training TimeGradPruner
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GradPrunerFine-Tuning Protocol=Full Fine-Tuning (FFT)2026.01 | 62.4 | 65.8 | 61.5 | 60.9 | |
| GradPrunerFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 66.9 | 64.4 | 61.8 | 60.7 | |
| LacoFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 71.5 | 65.7 | 61.1 | 61.4 | |
| LacoFine-Tuning Protocol=Full Fine-Tuning (FFT)2026.01 | 73.8 | 64.4 | 59.7 | 61.3 | |
| SATFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 74.8 | 82.9 | 102.3 | 99.4 | |
| SATFine-Tuning Protocol=Full Fine-Tuning (FFT)2026.01 | 75.5 | 79.4 | 98.9 | 103.6 | |
| LLMPrunerFine-Tuning Protocol=Full Fine-Tuning (FFT)2026.01 | 78.3 | 284.4 | 67.4 | 65.3 | |
| LLMPrunerFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 81.8 | 266.4 | 69.7 | 64.6 | |
| Dense ModelFine-Tuning Protocol=Full Fine-Tuning (FFT)2026.01 | 100 | 100 | 100 | 100 | |
| Dense ModelFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 100 | 100 | 100 | 100 | |
| APTFine-Tuning Protocol=Parameter-Efficient Fine Tuning (LoRA)2026.01 | 158 | 65.9 | 87.5 | 62.4 |