CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 14: Analysis, Interpretability, and Frontiers · part 107 of 111
Part 107 · Module 14: Analysis, Interpretability, and Frontiers

Lab: Instruments, circuits and a state-space layer

100 min read·13 Oct 2026
Module 14 · Lab · About 60 minutes · Needs: lessons 1 to 4 of this module, and the Module 13 lab

What you are building

TINYFORMER-LAB AFTER MODULE 14 analysis.py tasks with known answers hooks, ablation, patching ssm.py recall, SSMMixer train_mix, verdict v0.13.0 package unchanged; nothing in it edited test_m14_analysis 9 tests, no training m14_analysis run shortcut, circuit, SSM, base git tag v0.14.0
Two new files (blue) on an unchanged package (grey), pinned by nine tests (green) and measured by one run (amber).

Sol 452 at Aurora Station. Kofi Mensah asked what base has learnt. Four lessons built the instruments to answer him, each in its own cells, and left three things for the lab to settle.

Three induction tasks: the shared TASKS["induction"] puts every copy 16 back, and Lesson 1 caught a model passing it with "read 15 back". Lesson 2's make_varied gives each row its own gap, and two layers reach 0.999 with real previous-token and induction heads. Lesson 3's gap_task levelled at 0.68. Lesson 4 added make_recall. The package keeps the shared task as the shortcut example, Lesson 2's recipe as the induction task, and Lesson 4's recall.

Today they become tinyformer/analysis.py and tinyformer/ssm.py, nine tests, one experiment, tag v0.14.0.

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.