Lesson 2: Full fine-tuning and what it forgets

Module 12 · Lesson 2 · About 50 minutes · Needs: Lesson 1 (the evaluation, sft_example and the 448-character window); Module 6 (the base recipe)

The situation

Sol 440 at Aurora Station. Lesson 1 wrote Maya Lindqvist's evaluation and measured base: 0 of 200 cards right. Today you take the obvious next step and fine-tune every weight of base on the 1,000 training pairs.

Kofi Mensah asks for two numbers from every run, because Maya's bar has two gates: card accuracy on CARD_EVAL, and general validation loss no more than 0.02 above base's. The running example is log 5190 again: Omar Winters, the airlock motor, the comms mast, "pressure held steady".