Tokenizer design and KL-regularized fine-tuning

Discuss how you would design a tokenizer for a language model and the trade-offs involved. Then explain supervised fine-tuning with a KL penalty that keeps the model close to its starting point: why it helps and how it is computed.