Check out our new paper Safety Arithmetic at
https://arxiv.org/abs/2406.11801v1 👈
We introduce safety arithmetic, a test-time solution to bring safety back to your custom AI models. Recent studies showed LLMs are prone to elicit harm when fine-tuned or edited with new knowledge. Safety arithmetic can be solved by first removing harm direction in parameter space and then steering the latent… See the full description on the dataset page:
https://huggingface.co/datasets/SoftMINER-Group/NicheHazardQA.
