text-data

This paper introduces TAKE (Trajectory-Aware Knowledge Estimation), a text dataset distillation framework that uses influence functions and optimal transport to reduce datasets to as little as 0.1% of their original size while preserving downstream task fidelity.