arata1/dpo-qwen-cot-merged
The arata1/dpo-qwen-cot-merged model is a 4 billion parameter language model, fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using Direct Preference Optimization (DPO) via Unsloth. It is optimized to improve reasoning capabilities through Chain-of-Thought (CoT) and enhance structured response quality. This model is designed for applications requiring improved logical coherence and adherence to preferred output formats.
Loading preview...
Model Overview
The arata1/dpo-qwen-cot-merged model is a 4 billion parameter language model, fine-tuned from the Qwen/Qwen3-4B-Instruct-2507 base model. It leverages Direct Preference Optimization (DPO), implemented with the Unsloth library, to align its responses with preferred outputs. This model incorporates full-merged 16-bit weights, eliminating the need for adapter loading.
Key Capabilities & Optimization
- Enhanced Reasoning (Chain-of-Thought): The primary objective of this DPO fine-tuning was to improve the model's ability to generate coherent and logical reasoning steps.
- Improved Structured Responses: Optimized to produce higher quality, more structured outputs based on a preference dataset.
- Direct Use: As a fully merged model, it can be used directly with the
transformerslibrary without additional configuration for LoRA adapters.
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 during training. The LoRA configuration (r=8, alpha=16) was merged into the base model.
Licensing
This model follows the Apache-2.0 license of its base model, Qwen/Qwen3-4B-Instruct-2507. The training data, u-10bei/dpo-dataset-qwen-cot, adheres to its respective dataset card terms.