Model Overview

The qingy2024/Qwen2.5-Math-14B-Instruct-Preview is a 14.8 billion parameter instruction-tuned model developed by qingy2019. It is fine-tuned from the unsloth/qwen2.5-14b-instruct-bnb-4bit base model, utilizing the Unsloth framework for accelerated training. The model was specifically trained for 400 steps on the garage-bAInd/Open-Platypus dataset, focusing on enhancing its instruction-following and reasoning capabilities.

Key Capabilities & Performance

This model demonstrates proficiency in various reasoning and instruction-following tasks, as evidenced by its evaluation on the Open LLM Leaderboard. Notable scores include:

• IFEval (0-Shot): 60.66
• BBH (3-Shot): 47.02
• MATH Lvl 5 (4-Shot): 28.47
• MMLU-PRO (5-shot): 48.12

These results suggest its suitability for tasks requiring complex problem-solving and mathematical understanding.

When to Use This Model

• Mathematical Reasoning: Its specific fine-tuning and performance on MATH Lvl 5 indicate a strength in mathematical problem-solving.
• Instruction Following: The model is instruction-tuned, making it effective for general conversational AI and task execution based on prompts.
• Research and Development: Ideal for researchers and developers exploring efficient fine-tuning techniques with Unsloth and evaluating performance on reasoning benchmarks.