Core

Catastrophic Forgetting

When you fine-tune a model on new data, gradient updates overwrite the weights that encoded old skills, so the model gets better at the narrow new task and quietly worse at things it used to do well. It bites in practice when a model fine-tuned on a customer task loses general instruction-following. The fixes are replay data, parameter-efficient methods like LoRA, lower learning rates, regularizing toward the base, and always evaluating on a held-out general set before and after.

a free account unlocks the core curriculum tier · no card

Related concepts

Overfitting and RegularizationFree Golden Datasets and Eval SetsFree Bias-Variance TradeoffFree

Practise this in questions

AI Security, Privacy & GovernanceWhat is data/model poisoning, and how would you detect a backdoor introduced through fine-tuning or embedding data? →medium · Free RAG & Agent System DesignRetrieval fails on the customer's internal jargon. Fine-tune embeddings, add a reranker, or rewrite queries? →medium · Free LLM & GenAI FundamentalsA customer wants the model to 'know our docs.' Prompting, RAG, or fine-tuning: how do you choose? →medium · Free with sign-in LLM & GenAI FundamentalsWhen does fine-tuning beat few-shot prompting? How much data do you need, and how do you prevent regressions? →medium · Free with sign-in Behavioral & Customer ScenariosThe customer insists on fine-tuning when RAG clearly fits. They won't budge. Trusted advisor or vendor, what do you do? →medium · Free with sign-in LLM & GenAI FundamentalsLoRA vs full fine-tuning, and when does running a 4-bit quantized model on-prem make sense? →hard · Premium