How does training & evaluation work in CrewAI?

Question

I went through this post: How does training & evaluation actually work in CrewAI?

But I don’t think its that easy.

If just the prompts are injected in next run then why its called a training?

I see 2 .pkl files are saved training_data.pkl and usecase.pkl

Does it update some weights?

I would like to understand step by step of how training happens.

I assume we are not updating the weights of underlying LLM models as those might be big gpt models.