Model Overview

xiaodongguaAIGC/xdg-llama-3-8B is an 8 billion parameter language model built upon the Llama-3 architecture. It has undergone a comprehensive training regimen involving Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning from Human Feedback (RLHF), which includes both a reward model and Proximal Policy Optimization (PPO).

Key Capabilities

• Multifunctional: Excels in coding, general reasoning, and Chinese question-answering tasks.
• Safety Features: Incorporates safe refusal functions to prevent harmful or unethical responses.
• Training Methodology: Utilizes advanced alignment techniques (SFT, DPO, RLHF) for improved performance and safety.

Evaluation Insights

While the base Llama-3-8B model shows strong performance (MMLU 66.6, C-EVAL 49.5, C-MMLU 50.8), the xdg-llama-3-8B variant achieves MMLU 56.71, C-EVAL 42.83, and C-MMLU 45.04. This indicates a specialized focus or different optimization goals compared to the original Llama-3-8B.

Good For

• Applications requiring a balance of coding and reasoning abilities.
• Chatbots and conversational agents needing robust Chinese language support.
• Use cases where ethical and safe AI responses are critical, demonstrated by its refusal capabilities for inappropriate queries.