Llama 3 8B Instruct Ja

🚀 Llama 3 8B for Japanese

This repository contains a model that attempts to adapt Llama 3 for the Japanese language.

🚀 Quick Start

This repository hosts a model aiming to adapt Llama 3 for the Japanese language. Updated on April 23rd, it's recommended to download the latest version.

📄 License

It's under the Llama 3 License, allowing commercial use. However, please read the Llama 3 license carefully before use.

💻 Usage Examples

Basic Usage

If you want a quick try, use the demo. Another good option is Colab.

For local use, follow these steps:

First, install the libraries as follows:

pip install -U transformers accelerate

Then, run the following code:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("alfredplpl/Llama-3-8B-Instruct-Ja")
model = AutoModelForCausalLM.from_pretrained("alfredplpl/Llama-3-8B-Instruct-Ja", device_map="auto", torch_dtype=torch.bfloat16)

# Prepare the prompt
messages = [
    {
        'role': "system",
        'content': "あなたは日本語で回答するAIアシスタントです。"
    },
    {
        'role': "user",
        'content': "猫と犬、どっちが好き？"
    }
]
prompt=tokenizer.apply_chat_template(messages, tokenize=False)

# Run inference
input_ids = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **input_ids,
    max_new_tokens=128,
    do_sample=True,
    top_p=0.95,
    temperature=0.2,
    repetition_penalty=1.1,
    eos_token_id=[
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|eot_id|>")
    ],
)
print(tokenizer.decode(outputs[0]))

You should get a result like this:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

あなたは日本語で回答するAIアシスタントです。<|eot_id|><|start_header_id|>user<|end_header_id|>

猫と犬、どっちが好き？<|eot_id|><|start_header_id|>assistant<|end_header_id|>

猫と犬の両方を飼っているので、どちらも好きだ！<|eot_id|>

🔧 Technical Details

Training Data

• llm-jp/databricks-dolly-15k-ja
• cl-nagoya/auto-wiki-qa
• meta-llama/Meta-Llama-3-8B-Instruct

Training Method

We performed 1 epoch of LoRA-based instruction tuning on meta-llama/Meta-Llama-3-8B-Instruct using approximately 2.4 million training samples from cl-nagoya/auto-wiki-qa and then merged the LoRA. After that, we conducted 5 epochs of LoRA-based instruction tuning on the resulting model using llm-jp/databricks-dolly-15k-ja and merged the LoRA again. All these trainings were done using supervised learning.

Hardware

• NVIDIA A6000x2

Software

• transformers
• trl

Training Time

• 60 GPU hours