Nora Belrose

Bergson: An Open-Source Library for Scalable Data Attribution in Large Language Models

Bergson is a new open-source library engineered to advance research in data attribution for large language models. It addresses the significant engineering challenges in this field by providing scalable techniques, native support for on-disk gradient stores and multi-node distributed training, and open-source implementations of key methods like MAGIC, SOURCE, and TrackStar. This empowers researchers to more effectively explain model behavior by understanding the impact of training data.

Slowing Learning by Erasing Simple Features

Examining Two Hop Reasoning Through Information Content Scaling

Mechanistic Anomaly Detection for "Quirky" Language Models

Evaluating SAE interpretability without explanations

Binary Sparse Coding for Interpretability