CourseLarge Language Models: The Complete Course · Module 9: Fine-Tuning and Adaptation · part 64 of 103
Part 64 · Module 9: Fine-Tuning and Adaptation

Lesson 3: Train SmolLM2 two ways and score the test once

27 min read·13 Oct 2026
Module 9 · Lesson 3 · About 55 minutes · Needs: Module 9 Lessons 1 and 2 (the frozen plan, the bar and finetune/sft_train.jsonl); Module 1 Lesson 4 helps (label scoring)

The situation

It is Wednesday 29 July 2026, week 9. The plan is frozen, the bar is set, and Lesson 2 left 48 clean training rows. Tomasz wants to know which way to train: update every weight of SmolLM2-135M-Instruct, or train a small add-on. Joel wants to know what each costs. Maya wants one number: does the fine-tune clear the bar on the 12 test tickets?

You will follow one training row, T-1001, Ana Duarte's duplicate charge, from its answer tokens to the trained model, then score the test set once. Everything runs on a CPU.

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.