kikansha-Tomasu/sft-dpo-qwen-cot-merged

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Feb 11, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

kikansha-Tomasu/sft-dpo-qwen-cot-merged is a 4 billion parameter causal language model, fine-tuned from kikansha-Tomasu/Qwen3-4B-Instruct-2507-SFT using Direct Preference Optimization (DPO). This model specializes in improving reasoning capabilities through Chain-of-Thought (CoT) and enhancing structured response quality. It is optimized for tasks requiring aligned, high-quality outputs, particularly in reasoning and structured generation.

Loading preview...

Model Overview

This model, kikansha-Tomasu/sft-dpo-qwen-cot-merged, is a 4 billion parameter language model derived from kikansha-Tomasu/Qwen3-4B-Instruct-2507-SFT. It has been further fine-tuned using Direct Preference Optimization (DPO) via the Unsloth library, with its LoRA adapters merged into the base model for direct use.

Key Capabilities

  • Enhanced Reasoning: Optimized to improve Chain-of-Thought (CoT) reasoning, leading to more logical and structured problem-solving.
  • Improved Response Quality: Aligned through DPO to generate preferred outputs, focusing on high-quality and structured responses.
  • Direct Usage: Provided as a full-merged 16-bit model, eliminating the need for adapter loading and simplifying deployment with transformers.

Training Details

The model underwent 1 epoch of DPO training with a learning rate of 1e-07 and a beta value of 0.1. It utilized a maximum sequence length of 1024 during training. The preference dataset used for DPO was [u-10bei/dpo-dataset-qwen-cot].

Licensing

This model operates under the MIT License, consistent with its training data. Users are also required to adhere to the licensing terms of the original base model.