Catastrophic forgetting

Catastrophic forgetting is when you fine-tune a model on a new task and it loses general abilities it previously had. You taught it French legal text, and now it’s worse at math and coding.

It happens because the fine-tuning strongly pushes the weights toward the new distribution.

Mitigations:

• Mix in a bit of general data during fine-tuning.
• Use a low learning rate.
• Freeze some layers.