BrewSLM Academy · Track 2 — Hands-on: fine-tune a small model by hand

The hands-on track of the BrewSLM Academy — from a fresh Python environment to a fine-tuned model shipping behind an API, written by hand. Raw Trainer first for understanding, then TRL's SFTTrainer for production. QLoRA via bitsandbytes, sklearn + HF evaluate for real metrics, Pydantic for structured outputs, multi-turn chat SFT, and a project gallery of six SLM use cases to close. Honest by design — structured outputs need a two-number report; pick one project and ship it before you scope a second. Free, technical, zero-to-hero: https://brewslm.com/academy/hands-on/ Chapters: 0:00 Intro — Track 2 Hands-on 0:25 Set up the environment 0:49 Load a base model + tokenizer 1:18 Build a tiny SFT dataset 1:43 Tokenize & collate 2:14 A minimal LoRA fine-tune with the Trainer 2:45 Run it: read the logs 3:13 Evaluate by hand: run the gold set 3:42 Merge the adapter, run inference, ship 4:07 Capstone A: fine-tune end-to-end by hand 4:32 SFT with TRL's SFTTrainer (the 20-line version) 4:59 QLoRA hands-on with bitsandbytes 5:30 Real metrics with sklearn & HF evaluate 5:58 Structured outputs with pydantic 6:29 Multi-turn chat SFT 7:01 Project gallery: 6 SLM use cases 7:40 Recap — onward to Track 3