Follow the Gradient: Tracing Safety Degradation Back to Training Data
Open paper
Benchmark
Model
Rank
Results
arithmetic-reasoning-on-gsm8k
Llama 3.1 8B (SafeTulu)
–
Accuracy: 67.5
mathematical-reasoning-on-math-500
Llama 3.1 8B (SafeTulu)
–
Accuracy: 41.7
Rank counts only results with a code link.