tmaoshima/dpo-qwen-cot-merged

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 28, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

tmaoshima/dpo-qwen-cot-merged is a 4 billion parameter Qwen3-based causal language model, fine-tuned using Direct Preference Optimization (DPO) via Unsloth. This model is specifically optimized to improve reasoning capabilities through Chain-of-Thought (CoT) and enhance structured response quality. It leverages a preference dataset for alignment, making it suitable for tasks requiring coherent and well-reasoned outputs.

Loading preview...

Model Overview

This model, tmaoshima/dpo-qwen-cot-merged, is a 4 billion parameter language model based on Qwen/Qwen3-4B-Instruct-2507. It has been fine-tuned using Direct Preference Optimization (DPO) via the Unsloth library, with its 16-bit weights fully merged into the base model.

Key Capabilities & Optimization

  • Enhanced Reasoning: Optimized specifically to improve Chain-of-Thought (CoT) reasoning, leading to more logical and structured responses.
  • Improved Response Quality: DPO training aligns the model's outputs with preferred examples, enhancing the overall quality and coherence of generated text.
  • Direct Usage: As a fully merged model, it can be used directly with the transformers library without requiring adapter loading.

Training Details

The model underwent 1 epoch of DPO training with a learning rate of 2e-07 and a beta value of 0.02. It utilized a maximum sequence length of 512 tokens during training. The training data, u-10bei/dpo-dataset-qwen-cot, was instrumental in guiding the preference optimization process.

Licensing

Users must adhere to the MIT License as per the dataset terms and comply with the original base model's license terms.