merge-and-rebase
A research codebase for fine-tuning, merging and transporting task vectors between vision and language models, including rebasin methods that align independently trained networks before merging them.
merge-and-rebase supports fine-tuning, model merging, task-vector transport and evaluation across vision and text models from one configuration-driven codebase. A released Vision8 task-arithmetic example downloads its checkpoints automatically from the Hugging Face Hub and can be run directly from an evaluation script pointed at a configuration file.
Documentation covers the underlying concepts, task vectors, model bases, preparation and metrics, along with a configuration reference for merge, rebasin, post-merge and fine-tuning fields, a catalog of released checkpoints and manifests with their validation status, and a registry of the merge and rebasin methods themselves with their configuration parameters and source-level APIs. It installs with uv into a virtual environment, with optional dataset and development extras, and is Apache-2.0 licensed.