qwen3-4b-structured-output-base

wen3-4b-structured-output-v4plus5k-v1

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
• Datasets:
  • u-10bei/structured_data_with_cot_dataset_512_v4 (4,608 samples, system prompt enhanced)
  • daichira/structured-5k-mix-sft (filtered <1000 chars, ~2,500 samples)
• Method: QLoRA (4-bit)
• Max sequence length: 768
• Epochs: 1
• Learning rate: 1e-06
• LoRA: r=64, alpha=128

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/your-repo"

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:

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

Compliance: Users must comply with:

1. MIT License terms for both datasets (including copyright notice)
2. Base model's (Qwen3-4B-Instruct-2507) original terms of use