Catastrophic forgetting is what happens when a neural network trained on new data overwrites what it learned before. Because all of a model's knowledge lives in shared weights, updating those weights for a new task can erase the patterns that supported old tasks. A language model fine-tuned hard on legal contracts may become noticeably worse at everyday conversation, math, or following general instructions, even though nobody intended to remove those skills.
The problem is central to fine-tuning strategy. Aggressive training on a narrow dataset pulls weights far from their original values, and the narrower and longer the training, the more general capability is lost. Practical mitigations include lower learning rates and fewer epochs, mixing general instruction data back into the fine-tuning set, parameter-efficient methods like LoRA that leave the base weights frozen and learn small adapter layers instead, and regularization techniques that penalize drifting from the original weights. Often the best mitigation is avoiding fine-tuning altogether when retrieval can supply the knowledge at inference time.
At arosplatforms this shapes our default recommendation: use RAG for knowledge, fine-tune for behavior, and prefer LoRA-style adapters when fine-tuning is warranted. When we do train, we evaluate the resulting model on general-capability benchmarks as well as the target task, so a client never trades away broad competence for a narrow gain without knowing it.