Overview

alfredplpl/Llama-3-8B-Instruct-Ja is an 8 billion parameter instruction-tuned language model derived from Meta's Llama 3 architecture, specifically adapted for the Japanese language. The model has undergone a two-stage instruction tuning process to significantly improve its Japanese language capabilities.

Key Capabilities

• Enhanced Japanese Performance: The model was fine-tuned using approximately 2.4 million Japanese question-answering pairs from cl-nagoya/auto-wiki-qa and further refined with llm-jp/databricks-dolly-15k-ja.
• Instruction Following: It is instruction-tuned to respond effectively to user prompts, making it suitable for conversational AI and task-oriented applications.
• Commercial Use: The model adheres to the Llama 3 license, permitting commercial use.

Training Details

The training involved LoRA-based instruction tuning on an NVIDIA A6000x2 setup, accumulating 60 GPU hours. The process first merged LoRA adapters after training on cl-nagoya/auto-wiki-qa for one epoch, then repeated the merge after five epochs of training on llm-jp/databricks-dolly-15k-ja.

Good For

• Japanese AI Assistants: Ideal for building AI assistants that require strong Japanese language understanding and generation.
• Japanese Content Creation: Suitable for generating various forms of Japanese text, from creative writing to informative responses.
• Research and Development: Provides a solid base for further research and development in Japanese large language models.