catastrophic forgetting

Catastrophic forgetting occurs when neural networks lose previously learned knowledge while training on new data. **Mechanism**: Gradient updates for new task overwrite weights important for old tasks. Network doesn't distinguish between general knowledge and task-specific weights. **Symptoms**: Model excels at new task but fails at capabilities it previously had. Common when fine-tuning pretrained models on narrow domains. **Mitigation strategies**: Elastic Weight Consolidation (EWC) - penalize changes to important weights, memory replay - train on samples from previous tasks, progressive networks - add new capacity without overwriting, PEFT methods - freeze base model and train adapters, regularization techniques. **In LLM fine-tuning**: Aggressive learning rates cause forgetting, train on mixed data (old + new), use LoRA to preserve base capabilities. **Detection**: Evaluate on held-out benchmarks from original training distribution. **Practical advice**: Lower learning rates, shorter training, mix in instruction-following data, validate against base model capabilities regularly. Understanding forgetting dynamics is crucial for maintaining model quality during adaptation.