Model Overview

wanlige/li-14b-v0.4 is a 14.8 billion parameter language model developed by wanlige, leveraging the Model Stock merge method. It uses Qwen/Qwen2.5-14B-Instruct as its foundational base and integrates several specialized models, including Qwen2.5-Coder-14B, DeepSeek-R1-Distill-Qwen-14B, and others focused on text generation, chat, coding, and mathematics.

Key Capabilities

• Broad Task Performance: Designed for general text generation, chat assistance, and specialized tasks.
• Enhanced Reasoning: Incorporates models optimized for logic and mathematical problem-solving.
• Coding Proficiency: Includes components specifically for code generation and understanding.
• Instruction Following: Benefits from instruction-tuned and DPO-trained models for better adherence to prompts.

Performance Highlights

On the Open LLM Leaderboard, this model achieves an average score of 43.66. Specific metrics include:

• IFEval (0-Shot): 81.33
• BBH (3-Shot): 50.38
• MATH Lvl 5 (4-Shot): 55.74
• MMLU-PRO (5-shot): 46.30

Good For

• Developers seeking a versatile model for a wide array of NLP tasks.
• Applications requiring strong performance in both general language understanding and specialized domains like coding and math.
• Use cases where a merged model can provide a balanced blend of capabilities from multiple expert models.