mergoo

Why we included this project

Most teams with several fine-tuned checkpoints of the same base model would like the strengths of each without paying the cost of running them side by side. Mergoo handles that by routing Llama, Mistral, Phi3, or BERT checkpoints through a single network, and you can then fine-tune either the router alone or the whole merged model with the standard Hugging Face trainer APIs. It covers the two setups that actually come up in practice: merging fully fine-tuned experts into a mixture-of-experts stack, or putting a learned router on top of existing LoRA adapters. The library installs with pip, works with PEFT and SFTrainer, and ships notebooks for each merge method, which keeps the first prototype from turning into a routing-layer research project. If you are comparing model-composition strategies, configurable expert counts and router layers are built in.