Target:

How can we improve the interpretability of deep learning networks when applied to source code?

Project Objectives:

Develop a Web-Based Annotation and Visualization Tool

• Convert existing Tkinter annotation tools to a web application.
• Visualize concept clusters using interactive elements like word clouds, dendrograms, and sunburst charts.
• Integrate manual annotations and LLM-generated labels.

Generate Interpretability Datasets Using RAID

• Utilize the RAID (Rapid Automatic Interpretability Datasets) tool to create labeled datasets.
• Incorporate Tree-sitter for generating abstract syntax trees and apply B-I-O labeling.
• Visualize datasets similarly to Hugging Face's dataset viewer.

Ideally, we will combine these functionalities into a master app and/or repository. The aim is to have an app usable by both industry professionals and academics looking to improve their models or advance their research into deep learning and/or language models.

Figures

Web App Pipeline

RAID Workflow