Client: Dr. Ali Jannesari, Advisor: Arushi Sharma

We have extracted feature representations from code-trained large language models (LLMs) with the goal of explaining and interpreting these representations. By applying clustering algorithms, we have obtained clusters of latent concepts that represent various patterns learned by the models from the code.

This project focuses on auto-labeling code datasets to capture different code properties. We will use Abstract Syntax Tree (AST) tools like Tree-sitter, regular expressions, and LLM-generated labels to automatically annotate these datasets. Once the datasets are labeled, we will evaluate the concepts learned by the LLMs by measuring their alignment with the auto-labeled datasets. This evaluation will help determine how well the machine-learned concepts correspond to human-defined code properties, enhancing the interpretability of the models.