qwen3-4b-structevalt-lora-nobuta-daichira

qwen3-4b-structured-output-lora DPO qwen3-4b-structured-output-lora daichira/structured-5k-mix-sft

This repository provides a LoRA adapter fine-tuned from NobutaMN/dpo-qwen-cot-merged using QLoRA (4-bit, Unsloth).

This repository contains LoRA adapter weights only. The base model must be loaded separately.

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV).

Loss is applied only to the final assistant output, while intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

• Base model: NobutaMN/dpo-qwen-cot-merged
• Method: QLoRA (4-bit)
• Max sequence length: 512
• Epochs: 1
• Learning rate: 1e-06
• LoRA: r=64, alpha=128

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "NobutaMN/dpo-qwen-cot-merged"
adapter = "NobutaMN/dpo-qwen-cot-merged"

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

Sources & Terms (IMPORTANT)

Training data: daichira/structured-5k-mix-sft

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.

Reproducibility Notes

• Environment: Google Colab Free (T4 GPU).
• Seed: 3407
• Batch size: per_device_train_batch_size=2, gradient_accumulation_steps=8 (effective batch size=16)
• Precision: fp16
• Sequence length: 512
• Masking: assistant-only loss; CoT masked by learning only after output markers (Output:, Final:, Answer:, Result:, Response:).