Overview

This model, shinich001/dpo-qwen-cot-merged, is a 4 billion parameter instruction-tuned variant of the Qwen3-4B-Instruct-2507 base model. It leverages Direct Preference Optimization (DPO) via the Unsloth library to align its responses with preferred outputs. The repository provides the full-merged 16-bit weights, eliminating the need for adapter loading.

Key Capabilities

• Enhanced Reasoning: Optimized to improve Chain-of-Thought (CoT) reasoning, leading to more structured and logical outputs.
• Improved Response Quality: Fine-tuned to produce higher quality, aligned responses based on a preference dataset.
• Direct Usage: As a merged model, it can be used directly with the transformers library without additional configuration.

Training Details

The model was trained for 1 epoch with a learning rate of 1e-07 and a beta value of 0.1. It utilized a maximum sequence length of 1024 and incorporated LoRA (r=8, alpha=16) which has been merged into the base weights. The training data used was u-10bei/dpo-dataset-qwen-cot.

License

This model is released under the MIT License, adhering to the terms of its training dataset. Users must also comply with the original base model's license terms.