Fine-Tuning PolyCoder on the Devign Dataset for Vulnerability Detection
Jul 2025 - Oct 2025 (4 months)
Applied supervised fine-tuning (SFT) to adapt the PolyCoder 160M model to vulnerability detection using the Devign dataset, achieving strong performance in identifying insecure code patterns. Integrating reinforcement learning-based fine-tuning (RFT) strategies to further improve the model's precision in detecting code vulnerabilities, experimenting with custom training loops and reward functions.