Local AI Engineering with Ollama

Understanding What Happened

As said, a model is a huge pile of numbers (the weights). Training means slowly adjusting those numbers until the model does the task well. It happens in a loop that repeats over and over:

1. Show the model a batch of examples. For us, that's a schema plus a question, with the correct SQL as the answer.

2. Let it guess. The model produces its own answer using its current weights.

3. Measure how wrong it was. Comparing its guess to the correct answer gives a single number called the loss. High loss means very wrong, low loss means close.