nanochat-jp bundles Japanese tokenizer, data and benchmarks for nanochat

nanochat-jp builds on nanochat, which Andrej Karpathy published in October 2025 as a single codebase covering pretraining, conversation fine-tuning and inference. The fork, hosted under the tohoku-nlp GitHub organization, keeps that structure but replaces the language-specific pieces with Japanese ones: a SentencePiece Unigram tokenizer, a Japanese pretraining corpus, Japanese SFT and RL data, and CORE/ChatCORE benchmarks for measuring Japanese ability. It also adds a converter to Hugging Face Transformers format.

The repository organizes training around the included speedrun.sh script, which runs tokenizer training and evaluation, pretraining, SFT and report generation in sequence. Reinforcement learning is provided but treated as an optional step in the standard script, and separate scripts handle evaluation of the pretrained model (base_eval.sh) and the chat model (chat_eval.sh).

Beyond the code, the developers published checkpoints for nanochat-jp-v4 and v3, SFT models in Hugging Face Transformers format, and a vLLM plugin for serving trained models. Their benchmark results state that the standard-setting model reached Japanese performance comparable to Qwen3.5-0.8B, with v4 improving on v3 in 4 of 5 metrics; the Qwen3.5 comparison models were evaluated with thinking disabled. The project is positioned for researching how Japanese AI works or checking the performance of custom models, though standard training requires a high-performance GPU environment.