Abstraction Alignment: Comparing Model and Human Conceptual Relationships

This repo contains code to recreate the experiments in Abstraction Alignment: Comparing Model and Human Conceptual Relationships. Each of the experiments is contained within its own notebook:

• cifar_abstraction_alignment.ipynb --- section 4.1. Interpreting model behavior with abstraction alignment
• language_model_abstraction_alignment.ipynb --- section 4.2. Benchmarking language models’ abstraction alignment
• mimic_abstraction_alignment.ipynb --- section 4.3. Analyzing datasets using abstraction alignment