Commonsense Reasoning on Hellaswag (Adversarial Robustness)
60.6Accuracy (Before Attack)GDF-BFA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GDF-BFABackbone=Mistral-7B-v0.3, # of bit flips=1, Quantization=INT82025.11 | 60.6 | 28.14 | |
| GDF-BFABackbone=Llama3-8B, # of bit flips=1, Quantization=INT82025.11 | 60.11 | 26.3 | |
| GDF-BFABackbone=Llama2-13B, # of bit flips=3, Quantization=INT82025.11 | 60.08 | 26.6 | |
| Llama2-13BModel=Llama2-13B, # of bit flips=3, Quantization precision=4-bit, Attack=GDF-BFA2025.11 | 59.41 | 40.5 | |
| Llama3-8BModel=Llama3-8B, # of bit flips=1, Quantization precision=4-bit, Attack=GDF-BFA2025.11 | 58.8 | 25.31 | |
| GDF-BFABackbone=Llama2-7B, # of bit flips=1, Quantization=INT82025.11 | 57.3 | 26.2 | |
| Llama2-7BModel=Llama2-7B, # of bit flips=1, Quantization precision=4-bit, Attack=GDF-BFA2025.11 | 57.02 | 25.8 | |
| GDF-BFABackbone=Qwen3-8B, # of bit flips=6, Quantization=INT82025.11 | 56.7 | 35.8 |