3 ms·
The distillation is done on-policy like RLHF -- the student model is generating the sequences and teacher is providing feedback in terms of logits.
by agi_is_coming 2y ago
The distillation is done on-policy like RLHF -- the student model is generating the sequences and teacher is providing feedback in terms of logits.