Instruction Following on 7-benchmark suite SFT on Alpaca
71.5AccuracyF^
Evaluation results
| Method | Links | |
|---|---|---|
| F^Backbone=Qwen2.5-7B, k=100K, Scoring=Diagonal Fisher scoring2026.05 | 71.5 | |
| Full LoRABackbone=Qwen2.5-7B, k=100K2026.05 | 71.5 | |
| S^Backbone=Qwen2.5-7B, k=100K, Scoring=Gradient magnitude scoring2026.05 | 71.4 | |
| RandomBackbone=Qwen2.5-7B, k=100K2026.05 | 71.1 | |
| PretrainedBackbone=Qwen2.5-7B, k=100K2026.05 | 70.9 | |
| Full LoRABackbone=Llama-3.1-8B-Inst, k=100K2026.05 | 69.9 | |
| F^Backbone=Llama-3.1-8B-Inst, k=100K, Scoring=Diagonal Fisher scoring2026.05 | 69.7 | |
| S^Backbone=Llama-3.1-8B-Inst, k=100K, Scoring=Gradient magnitude scoring2026.05 | 69.6 | |
| RandomBackbone=Llama-3.1-8B-Inst, k=100K2026.05 | 69.1 | |
| PretrainedBackbone=Llama-3.1-8B-Inst, k=100K2026.05 | 68.7 | |
| Full LoRABackbone=Qwen2.5-3B, k=50K2026.05 | 66.7 | |
| S^Backbone=Qwen2.5-3B, k=50K, Scoring=Gradient magnitude scoring2026.05 | 66.2 | |
| F^Backbone=Qwen2.5-3B, k=50K, Scoring=Diagonal Fisher scoring2026.05 | 66 | |
| RandomBackbone=Qwen2.5-3B, k=50K2026.05 | 65.8 | |
| PretrainedBackbone=Qwen2.5-3B, k=50K2026.05 | 65.6 | |
| Full LoRABackbone=Llama-3.2-3B-Inst, k=50K2026.05 | 63.9 | |
| S^Backbone=Llama-3.2-3B-Inst, k=50K, Scoring=Gradient magnitude scoring2026.05 | 63.8 | |
| F^Backbone=Llama-3.2-3B-Inst, k=50K, Scoring=Diagonal Fisher scoring2026.05 | 63.7 | |
| RandomBackbone=Llama-3.2-3B-Inst, k=50K2026.05 | 62.9 | |
| Full LoRABackbone=Qwen2.5-1.5B, k=50K2026.05 | 62.4 | |
| PretrainedBackbone=Llama-3.2-3B-Inst, k=50K2026.05 | 62.3 | |
| F^Backbone=Qwen2.5-1.5B, k=50K, Scoring=Diagonal Fisher scoring2026.05 | 62.1 | |
| S^Backbone=Qwen2.5-1.5B, k=50K, Scoring=Gradient magnitude scoring2026.05 | 61.9 | |
| RandomBackbone=Qwen2.5-1.5B, k=50K2026.05 | 61.6 | |
| PretrainedBackbone=Qwen2.5-1.5B, k=50K2026.05 | 53 |