Llama-3-ELYZA-JP-8B is a large language model trained by ELYZA, Inc. Based on meta-llama/Meta-Llama-3-8B-Instruct, it has been enhanced for Japanese usage through additional pre-training and instruction tuning. (Built with Meta Llama3)

For more details, please refer to our blog post.

Quantization

We have prepared two quantized model options, GGUF and AWQ. This is the AutoAWQ model.

The following table shows the performance degradation due to quantization:

Model	ELYZA-tasks-100 GPT4 score
Llama-3-ELYZA-JP-8B	3.655
Llama-3-ELYZA-JP-8B-GGUF (Q4_K_M)	3.57
Llama-3-ELYZA-JP-8B-AWQ	3.39

Use with vLLM

Install vLLM:

pip install vllm

vLLM Offline Batched Inference

python

from vllm import LLM, SamplingParams

llm = LLM(model="elyza/Llama-3-ELYZA-JP-8B-AWQ", quantization="awq")
tokenizer = llm.get_tokenizer()

DEFAULT_SYSTEM_PROMPT = "あなたは誠実で優秀な日本人のアシスタントです。特に指示が無い場合は、常に日本語で回答してください。"
sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=1000)
messages_batch = [
    [
        {"role": "system", "content": DEFAULT_SYSTEM_PROMPT},
        {"role": "user", "content": "古代ギリシャを学ぶ上で知っておくべきポイントは？"}
    ],
    [
        {"role": "system", "content": DEFAULT_SYSTEM_PROMPT},
        {"role": "user", "content": "クマが海辺に行ってアザラシと友達になり、最終的には家に帰るというプロットの短編小説を書いてください。"}
    ]
]

prompts = [
    tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    for messages in messages_batch
]

outputs = llm.generate(prompts, sampling_params)

# Print the outputs.
for output in outputs:
    print(output.outputs[0].text)
    print("=" * 50)

vLLM OpenAI Compatible Server

Start the API server:

bash

python -m vllm.entrypoints.openai.api_server \
--model elyza/Llama-3-ELYZA-JP-8B-AWQ \
--port 8000 \
--host localhost \
--quantization awq

Call the API using curl:

bash

curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
  "model": "elyza/Llama-3-ELYZA-JP-8B-AWQ",
  "messages": [
    { "role": "system", "content": "あなたは誠実で優秀な日本人のアシスタントです。特に指示が無い場合は、常に日本語で回答してください。" },
    { "role": "user", "content": "古代ギリシャを学ぶ上で知っておくべきポイントは？" }
  ],
  "temperature": 0.6,
  "max_tokens": 1000,
  "stream": false
}'

Call the API using Python:

python

import openai

client = openai.OpenAI(
    base_url="http://localhost:8000/v1",
    api_key = "dummy_api_key"
)

completion = client.chat.completions.create(
    model="elyza/Llama-3-ELYZA-JP-8B-AWQ",
    messages=[
        {"role": "system", "content": "あなたは誠実で優秀な日本人のアシスタントです。特に指示が無い場合は、常に日本語で回答してください。"},
        {"role": "user", "content": "古代ギリシャを学ぶ上で知っておくべきポイントは？"}
    ]
)

Developers

Listed in alphabetical order.

License

Meta Llama 3 Community License

How to Cite

tex

@misc{elyzallama2024,
      title={elyza/Llama-3-ELYZA-JP-8B},
      url={https://huggingface.co/elyza/Llama-3-ELYZA-JP-8B},
      author={Masato Hirakawa and Shintaro Horie and Tomoaki Nakamura and Daisuke Oba and Sam Passaglia and Akira Sasaki},
      year={2024},
}

Citations

tex

@article{llama3modelcard,
    title={Llama 3 Model Card},
    author={AI@Meta},
    year={2024},
    url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md}
}