Applied AI·Fine-tuning and customisation
your fine-tune nails the target task and quietly got worse at everything else, which you only found out because you kept a general eval set.
Catastrophic forgetting
Draft summary, pending review
Fine-tuning degrading abilities the base model had, as new learning overwrites old. Mitigated by low learning rates, LoRA's light touch and mixing general data into the training set. Always eval a fine-tune on general tasks, not only the target task.