ML//Training//fine-tuning

Take a model that read the entire internet and teach it to do one thing properly. Take a pre-trained model, train further on task-specific data.

Take a model that read the entire internet and teach it to do one thing properly. Take a pre-trained model, train further on task-specific data.

Full fine-tuning updates all weights — expensive, risks catastrophic forgetting.

LoRA and adapters make it cheaper: freeze most weights, train a small delta.

The bridge between general knowledge and specific skill.