TEAtime4Code

About

Although current LLMs have shown strong performance in code generation tasks, their limitations around version awareness and contamination remain unresolved. Recent discussions around data provenance and legal risk have highlighted the need for more robust tools and processes. Code versioning, in particular, remains an open challenge, as current LLMs struggle with understanding differences, tracking context across versions, and assisting with refactorings in complex systems.

This project aims to explore two critical and under-addressed issues in the use of LLMs for code:

1. Contamination, the inadvertent leakage or memorization of licensed or sensitive code in LLMs
2. Versioning, the difficulty current models face in understanding and working with different versions of software libraries.

The project will evaluate and improve approaches using open source foundation models (Llama, Qwen and Appertus) to support the generation and evolution of secure, efficient and transparent code.