⚡ Activation

💭 in your head · nothing to type

Picture two teams. The first spends a week choosing a model and tweaking training settings. The second spends a week going through its examples: removing duplicates, fixing mixed-up answers, adding rare cases.

Both trained a model on the same hardware.

Whose model do you think will handle the real task better?

👁 Reflection

💭 in your head · nothing to type

Most likely you bet on the second. A model learns exactly what it was shown — mistakes in the examples included.

It's the same idea as in module 04: the agent does what you phrased, not what you meant. In training, your phrasing is the dataset. Every crooked example is a crooked spec, repeated many times.

Consider where in your work you already have "input → correct output" examples that someone has checked.

💡 Concept

The machine learning cycle

Every ML project, from an email classifier to fine-tuning an LLM, goes through the same seven steps:

1. Data collection — examples of the task, ideally from real life.
2. Preprocessing — cleaning, deduplication, a single format, labeling.
3. Model selection — the algorithm class and a specific model for the task.
4. Training — the model adjusts its weights to the examples.
5. Testing — checking on examples the model hasn't seen.
6. Tuning — fixing data and settings based on the test, then training again.