DistilT5: Knowledge Distillation for Automated Test Assertion Generation
This release presents DistilT5, a comprehensive implementation of knowledge distillation techniques that creates an efficient CodeT5-based model, specialized in generating test assertions from method bodies. As part of a Bachelor's Thesis at Delft University of Technology, this project addresses the computational challenges in automated test generation by producing a smaller, faster model while maintaining high-quality assertion predictions.
Key Features
- Complete distillation pipeline for CodeT5 models
- Custom training and evaluation components
- Performance measurement utilities
- Visualization tools for result analysis
- Logit decompression implementation for enhanced performance
This work contributes to the field of automated software testing by demonstrating how knowledge distillation can be effectively applied to create more efficient models for test assertion generation, potentially enabling wider adoption of these techniques in resource-constrained environments.