Overfitting occurs when a model learns the particular patterns of its training data so well that it struggles with new examples. A fine-tuned LLM may perform well on familiar inputs yet prove less reliable on unseen requests. This is an important trade-off when adapting a general model to an enterprise task. section 2.8.6 section 2.5
Watch for overfitting at two different points:
- During fine-tuning, compare performance with a baseline and evaluate the model on data it did not train on. Monitor results as the dataset changes; data quality and task complexity both matter. section 9.1.2
- During prompt design, avoid excessive direction. An overspecified prompt can cause an LLM to echo the prompt or leave too little room for novel output. This is overfitting to a prompt by analogy, not the same as fitting training data. section 6.7
For fine-tuning, keep a validation dataset separate from the training dataset used to update model parameters. Check validation performance while adjusting the learning rate, number of epochs, or other hyperparameters. Reserve a test dataset for measuring final accuracy. A longer context window can also increase model complexity and overfitting risk when training data is limited, so more context is not automatically better. section 9.3.1 section 2.8.6
A narrowly tailored response task
Suppose an LLM is fine-tuned to return emoji responses to user messages. Prepare curated training examples and separate validation examples for the same task. If training examples appear successful but validation results are poor, revisit the data and training settings rather than treating training performance as proof that the model will handle new messages. section 9.3.1
Aim for enough high-quality examples to represent the target task, then compare the adapted model with a baseline on new inputs. Prompt wording, dataset size, training settings, and context-window length each involve trade-offs; choose them by observed performance rather than by training results alone. section 9.1.2 section 9.3.1
