Tracking Universal Features Through Fine-Tuning and Model Merging
This work provides insights into feature dynamics in transfer learning for researchers, but it is incremental as it uses small-scale models and existing methods.
The study investigated how features change across models fine-tuned on different text domains, using a base Transformer model adapted to TinyStories and Lua, then merged via spherical linear interpolation to analyze feature stability and transformation.
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.