Prompt and example answer
Supervised fine-tuning (SFT) trains a model on example input-output pairs. A common objective increases the probability of desired output tokens conditioned on the prompt and earlier tokens.
Use a loss
A token assigned low probability incurs high negative log probability. Training aggregates such terms, but practical masking and loss conventions vary. A higher probability for the desired token reduces its negative log loss. The diagram shows one token rather than a whole sequence.
INTERACTIVE EXPLORATION
Score a desired token
Change the probability assigned to a desired token.
Examples are not guarantees
Fine-tuning can improve behavior on a distribution without ensuring correct behavior everywhere. Data quality, evaluation and the relation to pretraining matter.
Check your understanding
For a desired token assigned probability 0.5, is negative log probability positive or negative?
Positive: −ln(0.5) is about 0.693.