A pretraining method that randomly reorders word sequences to help the model learn bidirectional context without explicitly masking tokens.
Adhering to complex, structured, or constrained instructions
Multi-step reasoning, logic puzzles, mathematical problem-solving