ML//Training//parameter freezing

Choosing which model parameters to update during fine-tuning and which to lock (freeze)

Choosing which model parameters to update during fine-tuning and which to lock (freeze)

In standard fine-tuning and DPO: nothing is frozen by default — all parameters (attention + MLP) get updated.

In LoRA: base model frozen, small adapter matrices trained. Typically applied to W_Q, W_K, W_V, W_O — irónicamente las de attention, no MLP — because behavior changes most with fewest params there.

The tokenizer is almost never touched in fine-tuning — changing it means changing the embedding matrix, which is basically starting over.

Trainable vs frozen is a spectrum: full fine-tuning (0% frozen) → LoRA (~99% frozen) → prompt tuning (100% frozen, only input modified)