kotomamba-2.8B-v1.0

Kotomamba

The kotomamba model represents a cutting-edge approach in natural language processing (NLP), leveraging the innovative State Space Model mamba architecture. The kotomamba model comes in two distinct versions.

1. Bilingual Pre-training (Japanese and English): The first variant of the kotomamba model is pre-trained on a rich dataset(About 200B Token) comprising both Japanese and English texts.
2. Continual Pre-training (Mainly Japanese): The second variant of the kotomamba model takes a different approach, focusing exclusively on Japanese-centric data for its continual pre-training phase.

Kotomamba Model Index

This repository provides large language models developed by Kotoba Technologies, Tohoku University TohokuNLP group, and Tokyo Institute of Technology Okazaki Lab, Yokota Lab. Read our blog post or our technical paper (preprint coming soon) for more details!

Model Details

• Model type: Please refer to mamba technical paper for details on the model architecture.
• Language(s): Japanese English
• Library: kotomamba
• Tokenizer: kotomamba-2.8B uses llm-jp-tokenizer 100K and kotomamba-2.8B-CL uses GPT-NeoX Tokenizer.
• Contact:

Base Model Performance

Japanese version

Usage

git clone https://github.com/kotoba-tech/kotomamba and follow the repository's README installation section.

WARNING: huggingface transformers AutoModelForCausalLM doesn't support mamba model. So, please use kotomamba/benchmarks/benchmark_generation_mamba_simple.py

You can find the inference sample script in scripts/abci/inference/inference_sample.sh

Training Datasets

Pre-Training & Continual Pre-Training

The following datasets were used for training.

• Japanese Wikipedia
• Swallow Corpus
• SlimPajama

Risks and Limitations

The models released here are still in the early stages of our research and development and have not been tuned to ensure outputs align with human intent and safety considerations.

Acknowledgements

We thank Albert Gu and Tri Dao for releasing the original mamba model and implementation on GitHub.

Our project is supported by the ABCI Grand Challenge of the National Institute of Advanced Industrial Science and Technology.

License

Apache License Version 2.0, January 2004

Authors

Here are the team members:

• From Kotoba Technologies
  • Noriyuki Kojima
  • Jungo Kasai
  • Hiroto Kurita
  • Kazuki Fujii
• From TohokuNLP group at Tohoku University
  • Keisuke Sakaguchi
• From Tokyo Institute of Technologies
  • From Okazaki Laboratory, the following members:
    • Naoaki Okazaki
    • Sakae Mizuki
    • Hiroki Iida
    • Mengsay Loem
    • Shota Hirai
    • Kakeru Hattori
    • Masanari Ohi
  • From YOKOTA Laboratory, the following members:
    • Rio Yokota
    • Taishi Nakamura