---
language:
- ja
license: mit
library_name: transformers
pipeline_tag: text-generation
tags:
- mirei
- llama
- pretraining
base_model: iamtatsuki05/Llama-JP-0.5B-init
datasets:
- hotchpotch/fineweb-2-edu-japanese
---

# Llama-JP-0.5B-PT-stage1

English / [Japanese](README_JA.md)


## Overview
Llama-JP-0.5B-PT-stage1 continues from [iamtatsuki05/Llama-JP-0.5B-init](https://huggingface.co/iamtatsuki05/Llama-JP-0.5B-init) and is trained on [hotchpotch/fineweb-2-edu-japanese](https://huggingface.co/datasets/hotchpotch/fineweb-2-edu-japanese). The model observes approximately 10B tokens with 1,024-token context windows, providing a decoder-only Japanese backbone for downstream generative tasks.


- **[Hugging Face Collection](https://huggingface.co/collections/iamtatsuki05/mirei)**
- **[GitHub](https://github.com/iamtatsuki05/MIREI)**

![Consept](assets/concept.jpg)

## Usage
### Requirements
```
transformers>=4.51.0
accelerate>=1.6.0
sentencepiece>=0.2.0
flash-attn>=2.7.3
```

### Sample Code
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "iamtatsuki05/Llama-JP-0.5B-PT-stage1"
model_kwargs = {
  "torch_dtype": torch.bfloat16,
  "attn_implementation": "flash_attention_2",
  "device_map": "auto",
}
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)

prompt = "ちいかわのハチワレは"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.8,
    top_p=0.9,
    do_sample=True,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```


## Model Details
- **Base model:** [iamtatsuki05/Llama-JP-0.5B-init](https://huggingface.co/iamtatsuki05/Llama-JP-0.5B-init)
- **Architecture:** Llama
- **Maximum sequence length:** 8,192 tokens
- **Embedding dimension:** 1280
- **Tokenizer:** SentencePiece / vocabulary size 102,400
- **Positional encoding:** RoPE
- **Supported languages:** Japanese


## Model Series

The following checkpoints are initialized weights further pre-trained on [hotchpotch/fineweb-2-edu-japanese](https://huggingface.co/datasets/hotchpotch/fineweb-2-edu-japanese) for roughly 10B tokens with 1,024-token context lengths.

| ID                                                                                                            | Architecture | #Param. | #Param.<br>w/o Emb. |
|:-:|:-:|:-:|:-:|
| [iamtatsuki05/ModernBERT-JP-0.5B-PT-stage1](https://huggingface.co/iamtatsuki05/ModernBERT-JP-0.5B-PT-stage1) | ModernBERT   | 679M    | 548M                |
| [iamtatsuki05/Llama-JP-0.5B-PT-stage1](https://huggingface.co/iamtatsuki05/Llama-JP-0.5B-PT-stage1)<br>(this model)           | Llama        | 661M    | 530M                |


## Licence
This model is distributed under the [MIT License](https://opensource.org/license/mit/).


## How to Cite

```tex
@article{MIREI
  title={同一条件下における Encoder/Decoder アーキテクチャによる文埋め込みの性能分析},
  author={岡田 龍樹 and 杉本 徹},
  journal={言語処理学会第 32 回年次大会 (NLP2026)},
  year={2026}
}
```