Method weights training tokens by how much they predict answers

Original: Self-Privileged Fine-Tuning

Source: arxiv.org ↗