Catastrophic Forgetting: What Happens When LLMs Become Specialized

The Problem of Catastrophic Forgetting

Fine-tuning a model on specialized data may improve its performance in a specific domain while reducing some of its original general capabilities. This is known as catastrophic forgetting.

There are several ways to deal with this problem. The first option is simply to accept it. If the goal is to build a highly specialized model, losing some unrelated general capabilities may not matter. For example, a model designed for a specific professional task does not necessarily need to remain equally good at every general-purpose task.

The second approach is to include general-purpose data in the fine-tuning dataset. Instead of training only on specialized examples, we mix general and domain-specific data together. The proportion between the two types of data becomes an important consideration because it affects the balance between general and specialized capabilities.

The third approach is to use Adapters. Instead of updating the entire model, we can introduce a small number of trainable parameters while keeping the original model parameters largely frozen. LoRA is a widely used technique based on this idea.

These strategies reflect different choices. We can accept some loss of general capability, preserve it through a balanced training dataset, or adapt the model using techniques such as LoRA.

Fine-Tuning Is About Adaptation

The key idea behind fine-tuning is simple. We usually do not need to teach a Large Language Model everything again. The pretrained model already provides a strong foundation.

Fine-tuning allows us to adapt that foundation to a particular task, behavior, or domain. SFT teaches from examples, preference learning helps the model understand which outputs are preferred, and techniques such as LoRA make adaptation possible without retraining the entire model.

That is how a general-purpose LLM can gradually become a more specialized model for real-world applications.