Supervised fine-tuning
An example becomes a learning signal.
Choose the demonstrated response. Watch cross-entropy move probability toward it, one gradient update at a time.
Show the target
One of six outcomes is the demonstration. Its target probability is one; the other five targets are zero. This is one categorical prediction, not a complete language-model sequence.
Follow the gradient
The update subtracts learning rate times p minus the target from the logits. The colored motion shows that transfer; the bars and objective are computed from the updated logits.
Read the limits
SFT imitates a target. It does not collect environment reward or establish that the target is correct. Edit the target and initial logits to inspect a different example.
The equation behind the motion
L = −log p(y*) ∂L/∂zᵢ = pᵢ − 1[i = y*]
An exact six-outcome illustration. It omits neural-network and sequence-level training complexity.
Read the primary source ↗