qd2121_dl_final

Team: Spline Reticulator

Author: Thanh Do (qd2121) - only team member

Summary

Repository for the Deep Learning final project (ECE-GY 7123 / CS-GY 6953, Fall '25). This repo contains the materials for a project on Fine-Tuning Small Language Models for Targeted XML Fuzzing. The objective is to use a fine-tuned Small Language Model (SLM) as a learned fuzzer to generate XML/XSD inputs that maximize code coverage in the python-xmlschema library. The approach leverages Low-Rank Adaptation (LoRA) to fine-tune Qwen3-0.6B on the W3C XML Schema Test Suite, exploiting the model's "hallucination" behaviors as semantic mutations for effective fuzzing.

Key results

• Successfully fine-tuned Qwen3-0.6B with LoRA on W3C XML Schema Test Suite
• Generated XML/XSD samples show higher code coverage compared to zero-shot baselines
• Demonstrated that SLM "hallucinations" may serve as effective semantic mutations for structural fuzzing
• Model achieves lower perplexity on valid XML/XSD data after fine-tuning

Repository contents

• report/ — LaTeX source of the project report. Main file: report/acl_latex.tex.
• training_notebook.ipynb — training notebook used for fine-tuning the model with LoRA and generating test samples.
• baseline_samples/ — XML/XSD samples generated by the zero-shot baseline model (100 files).
• lora_samples/ — XML/XSD samples generated by the LoRA-tuned model (100 files).
• DL_Finals.pdf — prebuilt PDF report submitted to Gradescope.

High-level method

• Dataset: W3C XML Schema Test Suite (xsdtests). Contains valid and invalid XML/XSD files for training.
• Base model: Qwen/Qwen3-0.6B (a Small Language Model pre-trained on code).
• Training: LoRA fine-tuning with r=8, lora_alpha=16, trained with Trainer (Hugging Face Transformers) using adamw_torch, sequence length 1024, 15 epochs.
• Target Library: python-xmlschema version 4.2.0 — used for coverage evaluation.
• Approach: No-prompt training strategy (raw code, no instruction prompts) to bypass alignment and maximize generation of edge cases.

How to run the notebooks (local / GPU)

Requirements

• Python 3.10+ (recommended)
• CUDA-capable GPU
• The notebooks include dependency/setup cells (install Hugging Face libraries, PyTorch, coverage, xmlschema, etc.). Run those first.

Basic steps

1. Open training_notebook.ipynb in Google Colab or locally with GPU support.
2. Run the dependency/setup cells to install required packages (transformers, peft, datasets, torch, coverage, xmlschema).
3. Mount Google Drive (if using Colab) and configure paths to the W3C dataset.
4. Run training cells to fine-tune the model with LoRA.
5. Run generation cells to produce XML/XSD samples and evaluate code coverage on python-xmlschema.

CI: building the PDF report

A GitHub Actions workflow is provided at .github/workflows/build-report.yml. You can get the latest report by fetching the latest report-pdf artifact from the run page.

A prebuilt report file is included in the repo as DL_Finals.pdf which was submitted to Gradescope.

Please note that in order to view the Colab notebooks and access any shared resources, you'll need to log in to Google with your NYU account.