Overview

Mistral-7B-v0.3-Chinese-Chat is a 7.25 billion parameter instruction-tuned language model developed by Shenzhi Wang and team. It is built upon the mistralai/Mistral-7B-Instruct-v0.3 base model and has been full-parameter fine-tuned on a mixed Chinese-English dataset of approximately 100K preference pairs using the ORPO algorithm. This fine-tuning process has significantly enhanced its Chinese language abilities and reduced instances of mixed Chinese and English responses, a common issue in general-purpose models.

Key Capabilities

• Enhanced Bilingual Performance: Specifically optimized for both Chinese and English users, addressing common cross-lingual response issues.
• Diverse Functionality: Demonstrates strong performance in areas such as mathematics, roleplay, and tool use.
• Instruction Following: Instruction-tuned to follow user commands effectively across various tasks.
• GGUF Versions Available: Official q4, q8, and f16 GGUF quantized versions are provided for efficient deployment.

Good For

• Bilingual Chatbots: Ideal for applications requiring robust conversational abilities in both Chinese and English.
• Role-playing Scenarios: Capable of engaging in detailed role-play interactions.
• Mathematical Problem Solving: Shows good performance in solving mathematical problems.
• Tool Use/Function Calling: Designed to effectively utilize external tools based on instructions.
• Developers seeking a Mistral-based model with strong Chinese support.