Truly Open Japanese LLM Development

Fully Open,From Data to Models

We want Japanese LLMs to feel less like a finished product behind a wall and more like a workshop with the lights on. Bring a dataset, clean a script, test a model, write down what breaks, or suggest a better path. Small contributions should have a place to land and a clear way to become part of the next LLM build.

Open Data

Share datasets openly so anyone can inspect, improve, and reuse the data behind Japanese LLMs.

Open Source

Keep training, inference, and tooling code public so contributors can learn from it and build on it.

Open Model

Release model weights and details so the community can test, compare, and improve them together.

Small Enough to Hack,Big Enough to Explore

This project focuses on making Japanese LLM development understandable and accessible. It is built for people who want to read the code, run the models, improve the data, and learn by building together.

Codebase

The codebase is built from scratch with PyTorch so the main pieces are visible and easy to change. It includes tokenizer training, pretraining, midtraining, posttraining, evaluation, and inference tools, so contributors can follow the full path from raw text to generated output. Unlike large projects spread across many systems, the essential parts of building an LLM fit into one compact repository, making it easier to learn the full process and try your own changes.

Models

We keep model sizes practical so more people can train, inspect, and modify them without needing a large budget. Smaller models make it easier to understand what is happening inside, change the details, and learn how each choice affects Japanese text generation.

Explore the Latest Datasets and Models

Synthetic-Textbook-jp

A synthetic Japanese dataset that rewrites noisy web text into textbook-style passages for elementary, junior-high, and high-school levels. Built with Gemma 4, it helps LLMs learn natural Japanese flow from school-like topics.

lambda-1-160m-base

An experimental 164.5M parameter Japanese text generation model built with a decoder-only Transformer. It is open for testing, feedback, and improvement.

Help Us Build Better Open Models

Your support helps cover compute, data preparation, evaluation, and infrastructure. It keeps the work open and helps the community build better Japanese LLMs together.

Donate