CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 7: Architecture Families and Design Choices · part 51 of 111
Part 51 · Module 7: Architecture Families and Design Choices

Lesson 2: Landmark designs, rebuilt and counted exactly

18 min read·13 Oct 2026
Module 7 · Lesson 2 · About 50 minutes · Needs: Lesson 1 of this module (the three wirings); Module 3, Lesson 5 (count_params); Module 4 (position kinds)

The situation

Sol 418 at Aurora Station. Lesson 1 settled which wiring suits which job. Now Kofi Mensah asks a sharper question: "Is the lab's base a real design, or something we made up? Show me where it sits among the published models, and prove each comparison with a count."

This lesson rebuilds the landmark transformers as presets of the package's ModelConfig, compares each with base, and checks every weight count against the published checkpoint. The running example is base's own config.

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.