Self-Instruct

Wang et al 2022. Seed instructions + LLM generates more. Alpaca dataset created this way. Diversity + quality trade-offs.

Advertisement

Distillation

Teacher = big model. Student = small model. Train on teacher outputs. Vicuna, Phi trained this way.

Advertisement

Evol-Instruct

Iteratively make instructions harder ('rewrite this task with more constraints'). WizardLM approach.