qwen3-4b-structured-output-v3

qwen3-4b-structured-output-v3

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 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: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Datasets:
  • u-10bei/structured_data_with_cot_dataset_512_v4
  • daichira/structured-5k-mix-sft
• Max sequence length: 512
• Epochs: 2
• Learning rate: 2e-06
• LoRA: r=64, alpha=128

System Prompt Enhancement

Based on manual inspection of inference results, the following output rules were added to the system prompt:

• No conversational preambles ("Here's", "Here is")
• No explanatory text after output
• Clean and minimal responses

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "kasumi-nakano/qwen3-4b-structured-output-v3"

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 datasets:

• u-10bei/structured_data_with_cot_dataset_512_v4 (MIT License)
• daichira/structured-5k-mix-sft (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.