qwen3-4b-structured-output-lora-v5

This repository provides a LoRA adapter (v5) 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.

Version: v5 — XML Error Removal + Epoch=2

This is v5 of the SFT training, focusing on data quality improvement and hyperparameter tuning. Based on v2's success (0.75074), we applied three improvements.

Changes from v2

Parameter	v2	v5	Rationale
Dataset	3,933 samples	3,869 samples	XML errors removed (64 samples)
MAX_SEQ_LEN	1024	1024	Same as v2
Epochs	1	2	Person E's success (0.76+ with L4)
Learning Rate	5e-6	5e-06	Same as v2

v5 Dataset Details

• Base: u-10bei/structured_data_with_cot_dataset_512_v2 (3,933 samples)
• Removed: 64 samples with XML validation errors
• Final: 3,869 samples (98.4% of v2)

XML errors were detected using xml.etree.ElementTree parser.

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV) for the StructEval-T benchmark.

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, Unsloth)
• Max sequence length: 1024
• Epochs: 2
• Learning rate: 5e-06
• Batch size: 2 (effective: 16)
• Gradient accumulation: 8
• LoRA: r=64, alpha=128
• CoT masking: enabled (loss on final output only)

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/qwen3-4b-structured-output-lora-v5"

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: v5: XML error removed (3,869 samples from 1-1_512_v2)

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.