Track 3 · Training & adapting models
Fine-tuning LLMs
Fine-tuning is how you change a model's behavior with examples instead of longer prompts. This course teaches when it is worth doing, how to shape the dataset, which adaptation method to use, and how to ship the result without forgetting what made the base model useful.
01
02
03
04
05
06
When to fine-tune LLMs
The decision boundary between prompting, RAG, tool use, and fine-tuning. Fine-tune behavior, not missing facts.
Prepare a fine-tuning dataset
How examples become training signal: instruction formats, coverage, quality filters, splits, and what to leave out.
Full fine-tuning vs LoRA, PEFT, and QLoRA
What actually changes in the model, why adapters are cheap, and why most teams start with parameter-efficient fine-tuning.
Run a fine-tuning job
The training loop in practice: base model choice, learning rate, batch size, checkpoints, validation loss, and eval gates.
Avoid catastrophic forgetting
Why a model can get better at your task and worse at everything else, and how mixed data and evals catch it early.
Deploy fine-tuned models
Serving adapters, merging weights, versioning datasets, rollback plans, cost tradeoffs, and ongoing refreshes.