---
license: apache-2.0
datasets:
- SkelterLabsInc/JaQuAD
language:
- ja
---

 MambaSan-130m-instruct 🐍

**MambaSan-instruct is the first chat Japanese language model based on a state-space model architecture (Mamba), 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](https://arxiv.org/pdf/2312.00752.pdf)) as well as their [model implementation](https://github.com/state-spaces/mamba).
This work was also inspired by heavenq's mamba-chat implementation in English.

Mamba-Chat is based on MambaSan-130m and was fine-tuned on 31,7k examples samples of the [SkelterLabsInc/JaQuAD](https://huggingface.co/datasets/SkelterLabsInc/JaQuAD) dataset. To learn more, you can:

- Take a look at the model on [Huggingface](https://huggingface.co/loiccabannes/MambaSan-130m-instruct) 🤗
- Talk to Mamba-Chat on [Google Colab](https://colab.research.google.com/drive/1oDM071iXTLxiuDMzQtZVgyNzCi22xupy?usp=sharing)


The Code used for pretraining and finetuning will soon be published on my github: https://github.com/lcabannes
<br>


## Citation

```
bibtex
@misc{lcabannes2024MambaSan-130m-instruct,
  title        = {MambaSan-130-instruct},
  author       = {Loïc Cabannes},
  year         = {2024},
  howpublished = {HuggingFace},
  url          = {https://huggingface.co/loiccabannes/MambaSan-130m-instruct/}
}
```