Skip to content
BENIBIX Learn

Catastrophic Forgetting

AI💬 LLMsLanguage Model Training

Also known as Catastrophic Interference, Forgetting After Fine-Tuning

A model learns a new task and loses skills it had.

Think of it like

It is like you learn a new song so well that the old one is gone.

Example

Jack fine-tunes a chat model on his shop's help tickets. It now answers tickets well, but it starts to fail simple sums it used to get right.

Why it matters

A tuned model can pass its new test and still fail the old ones. You have to check both.

Definition

Catastrophic forgetting is when a model learns a new task and gets worse at old ones. Training changes the same weights that held the old skills, so new learning can write over them. In language models it often shows up after tuning on a narrow set of data. A model tuned only on legal text may get worse at math or chat. Mixing some old data back in, or changing only a few weights, can lower the risk.
After tuning on one kind of text, that skill rises and the others fall.

Common mistake

Testing only the new task. Run the old tests too, or you will not see what was lost.

Key takeaways

  1. 1
    Same weights. New learning writes over old.
  2. 2
    Narrow data. Raises the risk.
  3. 3
    Mix and limit. Old data and fewer changed weights help.

Compare

Learn it properly