Overview

nbeerbower/KawaiiMahou-llama3-8B is an 8 billion parameter language model built upon the Meta Llama 3 architecture. It has been fine-tuned by flammenai using Direct Preference Optimization (DPO) on a Japanese dataset, aiming to enhance its performance and alignment with preferred responses in Japanese contexts. The model is configured to use the ChatML format for its conversational interactions.

Key Capabilities

• Japanese Language Optimization: Fine-tuned specifically with a Japanese DPO dataset to improve relevance and quality for Japanese text generation.
• ChatML Format: Designed to operate with the ChatML format, enabling structured and consistent chat interactions.
• Llama 3 Base: Benefits from the foundational capabilities and performance of the Meta Llama 3-8B model.

Training Details

The model was fine-tuned using an A100 GPU on Google Colab. The DPO training involved specific LoRA configurations (r=16, lora_alpha=16, lora_dropout=0.05) and training arguments including a learning rate of 5e-5 and 1000 maximum steps. It utilizes paged_adamw_32bit optimizer and bf16 precision. The base model's license is governed by the META LLAMA 3 COMMUNITY LICENSE AGREEMENT.