token-order-prediction

Improve LM performance with token order prediction (TOP) auxiliary loss using learning-to-rank instead of exact multi-token prediction, achieving gains across math, code, and NLP tasks

Independent GitHub snapshot · Automated classification · Original documentation governs