Lesson 2: Build training data you can trust
29 min read·13 Oct 2026
Module 9 · Lesson 2 · About 50 minutes · Needs: Module 4 (the frozen dev and test sets); Module 9 Lesson 1 (the eval plan and the prompted baseline)
The situation
It is Monday 27 July 2026, week 9. The fine-tuning trial needs examples, and 24 dev tickets is not many. Tomasz exports the support log: 35 tickets that Lena and Ravi closed between 4 May and 21 July, each with the label the agent chose.
Maya asks: "Can we train on these?" Some rows hold emails and invoice numbers, some labels are wrong, and the queue that became your test set on 8 June was answered and closed like any other, so the test tickets are in this export. You will follow one row, L-0518, from the raw export to the tokens the model learns from.