Data Attribution
Which training examples contribute to a model's behavior? We develop tools for investigating that connection and use attribution to study capabilities and guide interventions.
Bergson provides an open-source library for training-data attribution. It supports the study of how particular examples influence models and makes attribution methods available for further research.
Our applications include investigating the provenance of social-reasoning capabilities and mitigating emergent misalignment through data attribution. These questions connect the study of model behavior to choices about training data.