このモデルは、Meta AI の Llama 3.1 をベースに、日本語タスクでの性能を向上させるためにファインチューニングを行ったものです。ElyzaTasks-100 において、オープンソースモデルで「gpt-4o-mini を超えるスコア／日本最高峰のスコア」を達成しました。
This model is based on Meta AI's Llama 3.1, fine-tuned to enhance performance on Japanese language tasks. It achieved "scores surpassing gpt-4o-mini / the highest scores in Japan" among open-source models in ElyzaTasks-100.
法的通知 / Legal Notice
This model is subject to the Llama 3.1 Community License Agreement. For detailed information, please refer to the official Llama license page: Llama 3.1 License
このモデルは Llama 3.1 Community License Agreement に従います。詳細については、Llama の公式ライセンスページをご参照ください。
開発者と使用者は、これらの制限事項を認識し、責任ある使用を心がけることが重要です。詳細については、Llama 3.1のResponsible Use Guideを参照してください。
This model, being based on Llama 3.1, carries similar limitations and ethical considerations:
Unpredictable Outputs: Like all LLMs, this model's potential outputs cannot be predicted in advance. It may sometimes generate inaccurate, biased, or problematic responses.
Need for Safety Testing: Developers should perform safety testing and tuning tailored to their specific applications before deploying any applications using this model.
Multilingual Considerations: While this model supports multiple languages, use in non-supported languages is not recommended without implementing fine-tuning and system controls aligned with appropriate policies.
Risks as New Technology: This model represents new technology and, like any new technology, there are risks associated with its use. Testing to date may not have covered all scenarios.
Need for Continuous Improvement: Continuous improvement of the model is necessary through community feedback and reporting mechanisms.
It's crucial for developers and users to be aware of these limitations and strive for responsible use. For more information, please refer to the Llama 3.1 Responsible Use Guide.
[Model Data]
Training Dataset]
We extracted high-quality data from Japanese Wikipedia and FineWeb to create instruction data. Our innovative training approach allows for performance improvements across various languages and domains, making the model suitable for global use despite its focus on Japanese data.
We used a plain instruction tuning method to train the model on exemplary responses. This approach enhances the model's ability to understand and generate high-quality responses across various languages and contexts.
