CourseTransformers: From Attention to Trained, Served and Interpreted Models · Module 8: Pretraining · part 59 of 111
Part 59 · Module 8: Pretraining

Lesson 3: Cleaning the corpus before it trains the model

32 min read·13 Oct 2026
Module 8 · Lesson 3 · About 55 minutes · Needs: Lesson 1 of this module (split, corpus report, BPE, packing) and Module 6 (the proxy and train)

The situation

Sol 421. Lesson 1's corpus report left Kofi Mensah with a question: 18 of the 200 lookup answer lines appear word for word in training documents. "Is the lookup score contaminated, or is that chance?" Meanwhile Tariq Haddad has pulled back the relay archive, Earth's mirror of the station record, as extra text. It arrived full of copies and garbled transmissions.

This lesson cleans the archive and answers Kofi. The running example is lookup question 4, whose answer line "Answer: Ana Ibarra works in the machine shop." also sits in training document 4420.

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.