Research
Researchers propose SimpleLoRIF to trace safety degradation to fine-tuning data
The method flags fine-tuning examples whose gradients align with unsafe behavior, and was used to build SafeTulu, a Tulu 3 variant with the 20% most unsafe-aligned examples removed. Training on SafeTulu cut GCG attack success rate by up to 22% across Qwen, Llama and Olmo, and the release includes a…
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
Follow the Gradient: Tracing Safety Degradation Back to Training Data