MSakae/dpo-qwen-cot-merged

＜Qwen3-4B Structured-Output SFT + DPO Aligned Model＞

This repository provides a merged 16-bit model derived from Qwen/Qwen3-4B-Instruct-2507. The model was trained using Direct Preference Optimization (DPO) via the Unsloth library, starting from an SFT-initialized LoRA adapter.

Training pipeline: Base model → Structured-output SFT (LoRA) → DPO preference alignment → merged_16bit export

What is included

✅ Full merged 16-bit weights (no adapter loading required)
✅ Tokenizer files
✅ Model card (this README)

Training Pipeline

Step 1 — SFT Initialization (LoRA)

Before DPO, the base model was initialized with an SFT LoRA adapter to improve structured output behavior (JSON / YAML / XML / TOML / CSV style formatting).

SFT initialization adapter: MSakae/qwen3-4b-structured-output-lora_sample_try_L4

Step 2 — DPO Alignment

The SFT-initialized model was further optimized using DPO with a preference dataset.

DPO dataset: u-10bei/dpo-dataset-qwen-cot

DPO aims to improve:

• Preference alignment between chosen vs rejected responses
• Response consistency and selection quality
• Structured response quality under preference constraint

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: DPO (Direct Preference Optimization)
• Epochs: 1
• Learning rate: 1e-06
• Beta: 0.1
• Max sequence length: 1024

Usage

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

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "your_id/your-repo-name"

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)

• Preference dataset: u-10bei/dpo-dataset-qwen-cot
• Base model terms: Users must comply with the original base model’s license/terms of use.
• Dataset terms: Users must comply with the dataset license/terms (including any required notices).