Llama3-Chinese-Chat is a fine-tuned version of Meta-Llama-3-8B-Instruct, specifically optimized for Chinese language interactions. It aims to address issues like "Chinese questions with English answers" and mixed language responses, offering enhanced capabilities in roleplay, function calling, and math for both Chinese and English users.

How It Works

This model is fine-tuned using ORPO (Odds Ratio Preference Optimization), a method that refines the model's preference for desired responses. It builds upon the Llama-3-8B-Instruct base model, leveraging a significantly larger dataset (up to 100K preference pairs in v2.1) to improve performance across various conversational tasks. The training framework used is LLaMA-Factory.

Quick Start & Requirements

• Ollama: ollama run wangshenzhi/llama3-8b-chinese-chat-ollama-q4 (for q4_0 GGUF)
• Hugging Face: Download GGUF versions (q4_0, q8_0, f16) from shenzhi-wang/Llama3-8B-Chinese-Chat-GGUF-* repositories.
• Transformers: Use transformers library with model_id = "shenzhi-wang/Llama3-8B-Chinese-Chat".
• Dependencies: Python, transformers, torch, llama-cpp-python (for GGUF). GPU recommended for optimal performance.

Highlighted Details

• v2.1 offers a 5x larger dataset than v1, with significant improvements in roleplay, function calling, and math.
• Demonstrates strong performance in roleplaying various personas (e.g., Taylor Swift, Jay Chou, Shakespeare) and handling complex prompts.
• Supports function calling with examples provided for internet_search and send_email.
• Includes examples of solving math problems and engaging in "Ruozhiba" style humor.

Maintenance & Community

• Developed by Shenzhi Wang (王慎执) and Yaowei Zheng (郑耀威).
• Model versions (v1, v2, v2.1) are available, with v2.1 being the latest.
• Training dataset for v2.1 is planned for release.

Licensing & Compatibility

• License: Llama-3 License.
• Compatible with commercial use under the terms of the Llama-3 License.

Limitations & Caveats

• The model's identity is not fine-tuned, leading to potentially random responses for identity-related questions.
• While v2.1 aims to reduce English word inclusion, it may still occur.