CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 11: Inference and Serving · part 79 of 111
Part 79 · Module 11: Inference and Serving

Lesson 2: Choosing a decoder, and fencing in what it writes

30 min read·13 Oct 2026
Module 11 · Lesson 2 · About 60 minutes · Needs: the shared data and base recipe (course-guide/06, blocks 1 to 11 and 15 to 19); Module 5's noise-floor habit; Module 10, Lesson 3 is helpful but not required

The situation

Sol 433 at Aurora Station. The Earth link is down for three sols, so base is the only model the station has. Maya Lindqvist has a backlog of maintenance logs that need cards, and Tariq Haddad's relay queue rejects anything not in the card's key: value format.

The model is fixed. What you can still choose is the decoder: the rule that turns each step's scores into one token. The running example is log 5190: Omar Winters rerouted the airlock motor in the comms mast.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.