ML//training//parameter freezing

Parameter freezing chooses which model parameters are updated during fine-tuning and which are locked (frozen). Trainable versus frozen is a spectrum:

Parameter freezing chooses which model parameters are updated during fine-tuning and which are locked (frozen). Trainable versus frozen is a spectrum:

1Full fine-tuning0% frozen2LoRAabout 99% frozen3Prompt tuning100% frozen, only the input changes

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

In LoRA the base model is frozen and small adapter matrices are trained, typically on WQW_QWQ​, WKW_KWK​, WVW_VWV​ and WOW_OWO​ (ironically the attention matrices rather than the MLP), because behavior changes most with the fewest parameters there.

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