qwen3-4b-structured-qlora-stage2-v1

qwen3-4b-structured-output-qlora-lora

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) and to reduce extraneous text.

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

Training Configuration

• Base model (card): Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Max sequence length: 1024
• Epochs: 1
• Learning rate: 1e-05
• LoRA: r=32, alpha=64

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/your-lora-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 datasets:

• daichira/structured-hard-sft-4k
• daichira/structured-5k-mix-sft
• daichira/structured-3k-mix-sft Dataset license(s): Please check each dataset card on Hugging Face and comply with their terms. Compliance: Users must comply with each dataset license and the base model's original terms of use.