Tracking Universal Features Through Fine-Tuning and Model Merging

Does a Python exception detection feature survive being adapted through Lua fine-tuning and diluted through slerp merging?

We extract language model features and track them across model adaptations, then explore high-level feature evolution dynamics and case studies.

Abstract

We study how features emerge, disappear, and persist across models fine-tuned on different domains of text. More specifically, we start from a base one-layer Transformer language model that is trained on a combination of the BabyLM corpus, and a collection of Python code from The Stack. This base model is adapted to two new domains of text: TinyStories, and the Lua programming language, respectively; and then these two models are merged using these two models using spherical linear interpolation. Our exploration aims to provide deeper insights into the stability and transformation of features across typical transfer-learning scenarios using small-scale models and sparse auto-encoders.