ML//training//fine-tuning

Fine-tuning is the further training of a pre-trained model on task-specific data, so that a model that read the entire internet learns to do one thing properly. It is the bridge between general knowledge and specific skill.

Fine-tuning is the further training of a pre-trained model on task-specific data, so that a model that read the entire internet learns to do one thing properly. It is the bridge between general knowledge and specific skill.

Full fine-tuning updates all weights, which is expensive and risks catastrophic forgetting. LoRA and adapters make it cheaper: they freeze most weights and train a small delta.

It is the one way among its neighbours that changes the model itself. Retrieval (RAG), stored notes (agent memory) and longer thinking (test-time compute) all change what the model sees or how long it works, leaving the weights as they were; fine-tuning bakes behaviour into the weights, which is why it suits a style, a format or a skill and suits poorly a fact that changes next month. Doing it repeatedly as new data arrives is the problem of continual learning.