README.md

---
language:
- ja
tags:
- japanese-stablelm
- causal-lm
pipeline_tag: text-generation
datasets:
- wikipedia
- CulturaX
license:
- other
extra_gated_prompt: >-
  By clicking "Agree", you agree to the [License Agreement](https://huggingface.co/stabilityai/japanese-stable-diffusion-xl/blob/main/LICENSE.md) and acknowledge Stability AI's [Privacy Policy](https://stability.ai/privacy-policy).
extra_gated_fields:
  Name: text
  Email: text
  Country: country
  Organization or Affiliation: text
  Receive email updates and promotions on Stability AI products, services, and research?:
    type: select
    options: 
      - Yes
      - No
---

# Japanese Stable LM 2 Instruct 1.6B

![A beautiful anime-like hummingbird flying with the text "Japanese Stable LM 2" below it, with a lofi anime landscape of Mount Fuji forming the outline of the text "Japanese Stable LM 2"](./japanese-stablelm-bird.png)

> A beautiful anime-like hummingbird flying with the text "Japanese Stable LM 2" below it, with a lofi anime landscape of Mount Fuji forming the outline of the text "Japanese Stable LM 2" — [Stable Diffusion 3](https://stability.ai/news/stable-diffusion-3)

Please note: For commercial use, please refer to [https://stability.ai/license](https://stability.ai/license)

## Model Description

`Japanese Stable LM 2 Instruct 1.6B` is a 1.6B-parameter decoder-only language model based on [Stable LM 2 1.6B](https://huggingface.co/stabilityai/japanese-stablelm-2-base-1_6b) that has been fine-tuned on a diverse collection of Japanese data, with the intent of maximizing downstream performance on Japanese language tasks.

## Usage

Japanese Stable LM 2 Instruct 1.6B uses the following instruction format:

```
<|user|>
「情けは人のためならず」ということわざの意味を小学生でも分かるように教えてください。<|endoftext|>
<|assistant|>
「情けは人のためならず」とは、優しいことをしてあげると、いつかそれが自分に返ってくるという意味のことわざです。<|endoftext|>
```

This format is also available through the tokenizer's `apply_chat_template` method:

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "stabilityai/japanese-stablelm-2-instruct-1_6b"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

# The next line may need to be modified depending on the environment
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True, 
    device_map="auto",
    trust_remote_code=True,
)

prompt = [
    {"role": "system", "content": "あなたは役立つアシスタントです。"},
    {"role": "user", "content": "「情けは人のためならず」ということわざの意味を小学生でも分かるように教えてください。"},
]
inputs = tokenizer.apply_chat_template(
    prompt,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

# this is for reproducibility.
# feel free to change to get different result
seed = 23
torch.manual_seed(seed)

tokens = model.generate(
    inputs,
    max_new_tokens=128,
    temperature=0.99,
    top_p=0.95,
    do_sample=True,
)

out = tokenizer.decode(tokens[0], skip_special_tokens=False)
print(out)
```

We suggest playing with different generation config (`top_p`, `repetition_penalty` etc) to find the best setup for your tasks. For example, use higher temperature for roleplay task, lower temperature for reasoning.

## Model Details

* **Model type**: `Japanese Stable LM 2 Instruct 1.6B` models are auto-regressive language models based on the transformer decoder architecture.
* **Language(s)**: Japanese
* **License**: See the [LICENSE file](https://huggingface.co/stabilityai/japanese-stablelm-2-instruct-1_6b/blob/main/LICENSE.md). 
* **Commercial License**: to use this model commercially, please refer to [https://stability.ai/license](https://stability.ai/license)
* **Contact**: For questions and comments about the model, please join [Stable Community Japan](https://discord.gg/StableJP). For future announcements / information about Stability AI models, research, and events, please follow [@StabilityAI_JP](https://twitter.com/StabilityAI_JP).

## Model Architecture

The model is a decoder-only transformer similar to the LLaMA ([Touvron et al., 2023](https://arxiv.org/abs/2307.09288)) architecture with the following modifications:

| Parameters     | Hidden Size | Layers | Heads | Sequence Length |
|----------------|-------------|--------|-------|-----------------|
| 1,644,417,024  | 2048        | 24     | 32    | 4096            |

* **Position Embeddings**: Rotary Position Embeddings ([Su et al., 2021](https://arxiv.org/abs/2104.09864)) applied to the first 25% of head embedding dimensions for improved throughput following [Black et al. (2022)](https://arxiv.org/pdf/2204.06745.pdf).
* **Normalization**: LayerNorm ([Ba et al., 2016](https://arxiv.org/abs/1607.06450)) with learned bias terms as opposed to RMSNorm ([Zhang & Sennrich, 2019](https://arxiv.org/abs/1910.07467)).
* **Biases**: We remove all bias terms from the feed-forward networks and multi-head self-attention layers, except for the biases of the query, key, and value projections ([Bai et al., 2023](https://arxiv.org/abs/2309.16609)).
* **Tokenizer**: We use Arcade100k, a BPE tokenizer extended from OpenAI's [`tiktoken.cl100k_base`](https://github.com/openai/tiktoken). We split digits into individual tokens following findings by [Liu & Low (2023)](https://arxiv.org/abs/2305.14201).


## Training Dataset

The following datasets were used for the instruction training.

- [jaster-v1.1.0](https://github.com/llm-jp/llm-jp-eval/blob/bbc03c655a93b244b6951f9549aad7dbf523508a/DATASET.md#jaster)
- [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF-%E5%85%AC%E9%96%8B/)
- [FreedomIntelligence/alpaca-gpt4-japanese](https://huggingface.co/datasets/FreedomIntelligence/alpaca-gpt4-japanese)
- [augmxnt/ultra-orca-boros-en-ja-v1](https://huggingface.co/datasets/augmxnt/ultra-orca-boros-en-ja-v1)

## Use and Limitations

### Intended Use

The model is intended to be used by all individuals as a foundation for application-specific fine-tuning without strict limitations on commercial use. For commercial use, please refer to https://stability.ai/license.


### Limitations and Bias

The pre-training dataset may have contained offensive or inappropriate content even after applying data cleansing filters which can be reflected in the model generated text. We recommend users exercise reasonable caution when using these models in production systems. Do not use the model for any applications that may cause harm or distress to individuals or groups.


## Authors

This model was developed by the Research & Development team at Stability AI Japan, and the development was led by Meng Lee (@leemeng) and Naoki Orii (@mrorii). The members of the team are as follows:

- [Meng Lee](https://huggingface.co/leemeng)
- [Naoki Orii](https://huggingface.co/mrorii)
- [Paul McCann](https://huggingface.co/polm-stability)
- [Yusuke Shibui](https://huggingface.co/cvusk)
- [Fujiki Nakamura](https://huggingface.co/fujiki)
- [Duy Phung](https://huggingface.co/pvduy)
- Maksym Zhuravinskyi
- Dakota Mahan
- [Jerry Chi](https://jerrychi.com)


## How to cite
```
@misc{JapaneseStableLM2Instruct1.6B, 
      url={[https://huggingface.co/stabilityai/japanese-stablelm-2-instruct-1_6b](https://huggingface.co/stabilityai/japanese-stablelm-instruct-2-1_6b)}, 
      title={Japanese Stable LM 2 Instruct 1.6B},
      author={Lee, Meng and Nakamura, Fujiki and McCann, Paul and Orii, Naoki and Shibui, Yusuke and Phung, Duy and Zhuravinskyi, Maksym and Mahan, Dakota and Chi, Jerry}
}
```