This model is a fine-tuned version of Qwen/Qwen3-4B-Instruct-2507 using Direct Preference Optimization (DPO) via the Unsloth library. The basemodel was merged with an adapter ykawasaki/qwen3-4b-structured-output-lora-v12 beforehand.

This repository contains the full-merged 16-bit weights. No adapter loading is required.

Training Objective

This model has been optimized using DPO to align its responses with preferred outputs, focusing on improving reasoning (Chain-of-Thought) and structured response quality based on the provided preference dataset.

Training Configuration

Base model: Qwen/Qwen3-4B-Instruct-2507
Adapter: ykawasaki/qwen3-4b-structured-output-lora-v12
Method: DPO (Direct Preference Optimization)
Epochs: 3
Learning rate: 1e-07
Beta: 0.1
Max sequence length: 1024
LoRA Config: r=16, alpha=32 (merged into base)
LoRA dropout: 0.05
rsLoRA: True
DoRA: False
Per device batch size: 16
Gradient accumulation: 8
Learning rate scheduler: SchedulerType.COSINE

Usage

Since this is a merged model, you can use it directly with transformers.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ykawasaki/qwen3-4b-dpo-qwen-cot-merged-v8"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Test inference
prompt = "Your question here"
inputs = tokenizer.apply_chat_template([{"role": "user", "content": prompt}], tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

Sources & License (IMPORTANT)

Training Data: [u-10bei/dpo-dataset-qwen-cot]
License: MIT License. (As per dataset terms).
Compliance: Users must follow the original base model's license terms.