A JumpStart text model fine-tuned hard on lock-tender notes now fails ordinary English prompts it handled last week. The new domain overwrote the old weights. What happened, and what should they try?
Select an answer to reveal the explanation.
Short Explanation
The JumpStart text model forgot ordinary English after a hard lock-tender fine-tune. That is catastrophic forgetting, so lower the rate, mix general examples, or freeze early layers. Training longer on only the new notes overwrites more, and Fraud Detector or dropout-only miss the diagnosis.
Full Explanation
Catastrophic forgetting is when fine-tuning on a new domain overwrites earlier general weights. Mitigate it with a lower fine-tune rate, mixed general examples, or frozen early layers. Training longer on only the new set deepens the overwrite. Fraud Detector and dropout-only miss that diagnosis.