Improving language model behavior by training on a curated dataset

Excerpt

Our latest research finds we can improve language model behavior with respect to specific behavioral values by fine-tuning on a small, curated dataset.