/mamba-chat

Mamba-Chat: A chat LLM based on the state-space model architecture 🐍

Primary LanguagePythonApache License 2.0Apache-2.0

Mamba-Chat 🐍

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. To learn more, you can:


Run Mamba-Chat

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 (CLI chatbot):

python chat.py

Talk to Mamba-Chat (gradio app):

pip install gradio==4.8.0
python app.py --share

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

If you have a 24GB card (3090, 4090, etc.) you can use these settings:

python train_mamba.py --model state-spaces/mamba-2.8b --tokenizer EleutherAI/gpt-neox-20b --learning_rate 5e-5 --batch_size 1 --gradient_accumulation_steps 4 --optim paged_adamw_8bit --data_path ./data/ultrachat_small.jsonl --num_epochs 3

Citation

bibtex
@misc{haven2023mambachat,
  title        = {Mamba-Chat},
  author       = {Justus Mattern and Konstantin Hohr},
  year         = {2023},
  howpublished = {GitHub},
  url          = {https://github.com/havenhq/mamba-chat}
}