---
base_model: Qwen/Qwen3-4B-Instruct-2507
datasets:
- u-10bei/structured_data_with_cot_dataset_512_v2
language:
- en
license: apache-2.0
library_name: peft
pipeline_tag: text-generation
tags:
- qlora
- lora
- structured-output
---

# qwen3-4b-structured-output-lora-v5.5C

This repository provides a **LoRA adapter (v5.5C)** 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: v5.5C — Hyperparameter Tuning

This is v5.5C of the SFT training, focusing on **hyperparameter tuning** to address overfitting in v5.
v5 achieved 0.73981, which was lower than v2 (0.75074) due to Epoch=2 causing overfitting.

### Changes from v5

| Parameter | v5 | v5.5C | Rationale |
|---|---|---|---|
| Dataset | 3,869 samples | **3,869 samples** | Same (XML errors removed) |
| MAX_SEQ_LEN | 1024 | **1024** | Same |
| Epochs | 2 | **1** | Reduced to prevent overfitting |
| Learning Rate | 5e-6 | **1e-06** | Lower for more stable training |
| Warmup Ratio | 10% | **10%** | Increased to reduce early instability |

### v5.5C Key Improvements

1. **Epoch=1**: v5's Epoch=2 caused overfitting (train/loss increased at end)
2. **Learning Rate 5e-6**: More conservative learning to prevent overfitting
3. **Warmup Ratio 10%**: Longer warmup for training stability
4. **Data unchanged**: 3,869 samples with XML errors removed

### Score History

| Version | Data | Score | Notes |
|---|---|---|---|
| v2 | 3,933 | **0.75074** | Best score baseline |
| v5 | 3,869 | 0.73981 | Epoch=2 overfitting |
| v5.5C | 3,869 | (pending) | Hyperparam tuning |

## 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: 1
- Learning rate: 1e-06
- Warmup ratio: 10%
- Batch size: 2 (effective: 16)
- Gradient accumulation: 8
- LoRA: r=64, alpha=128
- CoT masking: enabled (loss on final output only)

## Usage

```python
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-v5.5C"

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_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.