A 6.8 billion parameter pre-trained model for Japanese language, based on EleutherAI's Mesh Transformer JAX, that has a similar model structure to their GPT-J-6B pre-trained model.

EleutherAIによるMesh Transformer JAXをコードベースとした、GPT-J-6Bに似たストラクチャと約68.7億パラメータを持つ日本語pre-trainedモデルです。

- We used T5Tokenizer and SentencePiece instead of GPT-2/3 tokenizer. Normalization done by SentencePiece is must for Japanese tokenizing as there are so much many more variations for common symbols than Western languages.
- Tokenizer has a vocabulary of 52,500 tokens and trained on Japanese Wikipedia dump as of 01 Aug 2021.
- The model fits within 16GB VRAM GPUs like P100 for inference up to 1688 context length. Full 2048 context length output requires 20GB VRAM or more (e.g. GTX3090/A5000).
- The model was trained with TPUv3-128 generously provided by Google TRC for about 4 weeks.

## Specifications

| Hyperparameter    | Value  |
|-------------------|--------|
| n_parameters      | 6,876,450,080 |
| n_layers          | 32 |
| d_model           | 4,096  |
| d_ff              | 16,384 |
| n_heads           | 16     |
| d_head            | 256    |
| n_ctx             | 2,048  |
| n_vocab           | 52,512 |
| position encoding | [Rotary position encodings (RoPE)](https://arxiv.org/abs/2104.09864) |
| RoPE dimensions   | 64 |