nnsohamnn/Qwen2.5-3B-ReTrace-OpenO1-Merged
Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:Dec 13, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

nnsohamnn/Qwen2.5-3B-ReTrace-OpenO1-Merged is a 3.09 billion parameter Qwen2.5 Transformer model, fine-tuned by nnsohamnn. It is specifically optimized for structured reasoning tasks, generating step-by-step thought processes and final answers using and tags. This model excels in multi-domain reasoning, including math, logic, and general problem-solving, making it suitable for applications requiring explicit, verifiable reasoning paths.

Loading preview...

Overview

This is a fully merged Qwen2.5-3B-Instruct model, fine-tuned by nnsohamnn using LoRA on 5,000 reasoning samples from the ReTrace and OpenO1-SFT datasets. The model is designed to produce structured reasoning, explicitly detailing its thought process within <Thought> tags before providing a final answer in <Output> tags. This approach enhances transparency and verifiability in problem-solving.

Key Capabilities

  • Structured Reasoning: Generates explicit step-by-step thought processes and final answers.
  • Multi-Domain Problem Solving: Trained on diverse reasoning examples covering math, logic, word problems, and general reasoning.
  • Production Ready: Provided as a fully merged FP16 model, requiring no adapter loading, with a 6GB size.
  • Efficient Training: Achieved a 49.2% reduction in training loss over 310 steps, utilizing Unsloth and HuggingFace Transformers.

Good For

  • Applications requiring transparent, verifiable reasoning outputs.
  • Tasks involving mathematical problem-solving and logical deduction.
  • Use cases where a smaller, efficient model with strong reasoning capabilities is preferred.