ZYH-LLM-Qwen2.5-14B-V3 Overview

ZYH-LLM-Qwen2.5-14B-V3 is the third iteration in the ZYH-LLM series by YOYO-AI, a 14.8 billion parameter model based on the Qwen2.5 architecture. This model is distinguished by its advanced model merging techniques, designed to create a highly unified and powerful foundation for further fine-tuning and development. It boasts an impressive 131072 token context length, enabling it to process and understand extensive inputs.

Key Capabilities

• Superior Instruction Following: As of February 25, 2025, it holds the highest IFEval (0-Shot) score (85.78) among 14B models, indicating exceptional ability to understand and execute complex instructions.
• Robust Reasoning: Demonstrates strong performance in reasoning tasks, with a BBH (3-Shot) score of 48.18 and MATH Lvl 5 (4-Shot) score of 52.72.
• Unified Architecture: Developed through a multi-stage model merging process, integrating various Qwen2.5-14B variants and other models like EVA-Qwen2.5-14B-v0.2, arcee-ai/Virtuoso-Small-v2, and Azure99/Blossom-V6-14B.

Good for

• Applications requiring high-fidelity instruction adherence and complex task execution.
• Use cases demanding strong reasoning capabilities across diverse domains.
• Developers looking for a powerful and unified 14B base model for further specialized fine-tuning or merging experiments.