Working Notes: my commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog.

extended notes from building LLMs from scratch

Tokenizers

• special tokens to give the model more context
  • <|begin|> , <|unk|> etc.
  • bos, eos, pad, etc
• byte pair encoding
  • tiktoken up to gpt3 had a 50k vocab
  • merges tokens based on frequency
• from wikipedia
  • byte level encoding works
• of course karpathy has his own implementation
  • bye
  • in rust
• byte latent transformer
  • by mike, etc.

• huggingface

— Kunal