When Can First-Order Models of Fine-Tuning Bound Forgetting?
Fine-tuning a language model on new data can make it forget facts that it should keep. We ask whether measurements taken at the start of a fine-tuning run can bound, for each protected fact, the probability that the run makes the model forget it. In LoRA fine-tuning with stochastic gradient descent on models from 0.6B to 14B parameters, a first-order response model estimated by finite-difference…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
