Mamba-Chat is the first chat language model based on a state-space model architecture, not a transformer.
The model is based on Albert Gu's and Tri Dao's work Mamba: Linear-Time Sequence Modeling with Selective State Spaces as well as their model implementation. This repository provides training / fine-tuning code for the model based on some modifications of the Huggingface Trainer class.
We provide code that lets you run inference on mamba-chat as well as our fine-tuning code. To get started, clone this repository and install its dependencies:
git clone https://github.com/havenhq/mamba-chat.git
cd mamba-chat
pip install -r requirements.txt
You can chat with mamba by