24 upvotes
New method traces safety loss in LLM fine-tuning
What is Follow the Gradient: Tracing Safety Degradation Back to Training Data?
Fine-tuning large language models on seemingly harmless instruction data can silently reduce their safety. The authors introduce SimpleLoRIF, a gradient-based tool that identifies which training examples cause the degradation.
Who it is for
AI safety teams and developers can use it to locate and prune risky data before it harms model behavior.
How to try it
Open the paper on HuggingFace: https://huggingface.co/papers/2610.37481