Definition

SFT trains a base model on ~10K to 100K prompt-response demonstrations. It teaches the chat template: role markers like <|start_header_id|> and <|eot_id|> are single tokens whose meaning the model learns only here.

  • Base model: continues the pattern (β€œWhat is the capital of Germany? …”). Instruct model: answers (β€œThe capital of France is Paris.”).
  • Because roles are just tokens, fake format tokens in user input can confuse them (misdirection, system prompt extraction).

Appears in