← Back to Journal

Introduction to LLMs in Python

OCTOBER 6, 2026 · Career · LLM

I finished DataCamp's Introduction to LLMs in Python course, next after Responsible AI Data Management on the Associate AI Engineer for Data Scientists track.

Introduction to LLMs in Python course completion badge

What I'm taking away is how many challenges come with using an LLM. Evaluation is how you find them before production does, and in production model performance can play a big part in whether the thing works at all.

The lifecycle and the fine-tuning plan

An LLM starts with pre-training, which produces a foundation model. Fine-tune that foundation model and you get the fine-tuned model. The fine-tuning itself followed one plan every time: set the training plan, build the trainer, train, check the loss, predict, save.

Two ways to fine-tune

Full fine-tuning updates every weight in the model. Partial fine-tuning updates only some of them. The course also covered transfer learning, which reuses what a model already learned for a new task, and n-shot learning, where you show the model a few examples (or one, or none) instead of retraining it.

Evaluation

This was the chapter I got the most out of. The Hugging Face evaluate library had a metric for each kind of task: perplexity for how well a model predicts text, BLEU and METEOR for translation, ROUGE for summaries, and exact match (EM) for questions with one right answer. A different metric tells you something different about the same model, so picking one is part of the evaluation.

The course also covered safeguarding. Toxicity checks whether a model produces harmful text, and regard checks whether it talks about different groups with different sentiment. Both tie back to the bias work in Responsible AI Data Management. A model can score fine on BLEU or ROUGE and still say things you wouldn't want to ship.

What's next

Working with Llama 3.

CareerLLM