barryallen16/Qwen2.5_7B_java_vul_final

Uploaded model

• Developed by: barryallen16
• License: apache-2.0
• Finetuned from model : unsloth/qwen2.5-7b-instruct-unsloth-bnb-4bit

This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Fine-tuned LLM system that detects vulnerabilities in Java code. Built the full cycle myself: dataset curation → synthetic augmentation → dedup → LoRA fine-tuning → evaluation across model sizes and learning rates.

Headline result: Qwen2.5-Coder-7B fine-tuned at lr 3e-4 reaches 80.39% accuracy on 102 held-out test cases, beating my own 14B run (78.43%) and a 4-bit Mistral-Nemo-Instruct baseline (77.45%).

Results

Model	LR	Test cases	Correct	Accuracy
Qwen2.5-Coder 7B	3e-4	102	82	80.39%
Qwen2.5-Coder 7B (100 max steps)	2e-4	102	81	79.41%
Qwen2.5-Coder 14B	2e-4	102	80	78.43%
Mistral-Nemo-Instruct-2407 (4-bit)	2e-4	102	79	77.45%

The 7B outperforming the 14B suggests the task rewards cleaner training data over raw parameter count — consistent with what I saw while deduplicating: a meaningful share of the merged corpora were near-exact duplicates.

Pipeline

raw datasets → standardize → clean → LLM synthetic augmentation
   → merge → sha256 dedup → preprocess → conversation-format SFT set
   → Unsloth LoRA fine-tune → 102-case eval harness

• Datasets moved through seven curated stages (~1 GB raw); final SFT set is conversation-formatted for Unsloth/SFTTrainer.
• Deduplication by SHA-256 of normalized lines (datasets/deduped/).
• Eval harness (model-evaluation/testcases_java_eval.ipynb) runs a fixed 102-case set against any checkpoint via one function call: evaluate_model(llm_name, test_cases).

Artifacts

• Datasets: HF datasets
• Source Code: Github repo