distilled counterfactual data

Automated generation of high-quality counterfactual data.

DISCO (DIStilled COunterfactual Data) is a method for automatically generating high-quality counterfactual data at scale. It uses a large general language model to generate phrasal perturbations, which are then filtered by a task-specific teacher model to distill high-quality counterfactual data. The method has been applied to natural language inference tasks, demonstrating improved robustness and generalization across distributions.