Term
Catastrophic Forgetting
Catastrophic forgetting describes the effect that a neural network loses previously acquired knowledge — wholly or partially — when it is trained further on new data; a key problem when fine-tuning language models.
Catastrophic forgetting — explained in more detail
Language models store their knowledge distributed across all weights. When the same weights are optimised further on new training data, gradients inevitably overwrite information the model previously handled well. Consequence: a model post-trained on medical text suddenly answers everyday questions worse or even loses instruction-following ability. The severity depends on data volume, learning rate, training steps, and the distance between original and new data. The term originated in 1990s research on classic neural networks and has become central in the LLM era because continued pretraining and fine-tuning can tip a model over.
Example / Practical context
Typical symptoms in practice: after fine-tuning on 10,000 support tickets the model delivers the desired tone — but answers maths problems significantly worse than before training. Or: a German-specific continued pretraining improves German output but degrades code responses. Countermeasures: PEFT instead of full fine-tuning (adapters, LoRA — the base model stays untouched), replay (mixing old data into the new training set), lower learning rates, shorter training runs, regular evaluations on a broad benchmark set.
Distinction from related concepts
Overfitting is related but different: the model learns the training data too precisely and generalises worse. Catastrophic forgetting refers specifically to knowledge from earlier training phases that disappears. Continual learning is the research field that addresses both — methods that allow models to be extended permanently without destroying earlier abilities.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarDPO (Direct Preference Optimization)
DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.
LexikonFine-Tuning Explained — When, How, With What
When fine-tuning pays off, which methods exist (SFT, DPO, LoRA, QLoRA), and what AMD vs. NVIDIA hardware actually means in practice.