Supervised fine-tuning

Choose the demonstrated response. Watch cross-entropy move probability toward it, one gradient update at a time.

Preparing the preview…

Made to be adapted.

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.

A six-outcome mathematical model with exact updates. No language model is being trained.

Export formats: PNG / HTML
Editable in Studio: Explanation, appearance, and model inputs
Access: Free with attribution. No account needed.
Reference: Read the source

Loading your workspace…

Supervised fine-tuning

Made to be adapted.

Another angle on the idea.

Group-relative optimization

Knowledge distillation

Supervised fine-tuning

Made to be adapted.

Another angle on the idea.

Group-relative optimization

Knowledge distillation