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 (paper) 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.
Mamba-Chat is based on Mamba-2.8B and was fine-tuned on 16,000 samples of the HuggingFaceH4/ultrachat_200k dataset. We used a single A100 (40GB) GPU for training.
To learn more, you can:
- Take a look at the model on Huggingface 🤗
- Join the Haven Community Discord 🧑🤝🧑
We provide code for testing and fine-tuning our model. Here's how to get started and what you can do with it:
Clone repository and install dependencies:
git clone https://github.com/havenhq/mamba-chat.git
cd mamba-chat
pip install -r requirements.txt
Talk to Mamba-Chat:
python chat.py
Fine-Tune Mamba (the base model) on a subset of the Ultrachat dataset:
python train_mamba.py --model state-spaces/mamba-2.8b --tokenizer EleutherAI/gpt-neox-20b --learning_rate 5e-5 --batch_size 4 --data_path ./data/ultrachat_small.jsonl --num_epochs 3