A pretrained model already knows useful patterns, but its predictions may not fit your task. Fine-tuning continues training that model on task-specific data, adjusting some or all of its learned representations. For a small dogs-versus-cats dataset, for example, an ImageNet-pretrained Xception network offers a starting point instead of learning visual features from scratch. section 8.3.2 section 8.3
For image classification, use this order:
- Add a classifier to the pretrained convolutional base, freeze the base, and train the new classifier first. Unfreezing immediately lets large errors from the randomly initialized classifier damage the pretrained representations. section 8.3.2
- Then unfreeze the base and train it jointly with the classifier. Keep BatchNormalization layers frozen. With a large base, consider unfreezing only its upper layers: lower layers encode more generic features, while upper layers are more specialized. section 8.3.2
- Compile with a small learning rate, monitor validation loss, save the best checkpoint, and evaluate that saved model on test data. section 8.3.2
Listing 8.31 Fine-tuning the model
model.compile( loss="binary_crossentropy", optimizer=keras.optimizers.Adam(learning_rate=1e-5), metrics=["accuracy"], ) callbacks = [ keras.callbacks.ModelCheckpoint( filepath="fine_tuning.keras", save_best_only=True, monitor="val_loss", ) ] history = model.fit( augmented_train_dataset, epochs=30, validation_data=validation_dataset, callbacks=callbacks, )
The Xception training listing uses Adam at 1e-5, augmented training images, validation data, and a checkpoint selected by validation loss. section 8.3.2
The same idea applies beyond images. For IMDb sentiment classification, a pretrained RoBERTa backbone gains a binary classification head. Its first token’s output supplies one representation for the whole review; the resulting classifier is trained and evaluated on labeled reviews. Match the tokenizer’s pretrained sequence format, including start, end, and padding tokens, before training. In this example, one epoch reaches about 93% test accuracy, at a higher computational cost than a simpler bigram classifier. section 15.4.3 section 15.4.4
For a language model, instruction fine-tuning uses instruction–response pairs as further training data. Consistent markers distinguish the instruction from the response, and training can still use next-token prediction. This nudges output toward desired responses; it does not relearn language from zero. Handwritten responses provide supervised fine-tuning, but collecting them is slow and expensive. Reinforcement learning with human feedback builds on that baseline using ranked responses and a reward model. Fine-tuning can improve instruction following, but it is not a cure for hallucinations. section 16.3.2 section 16.4.1 section 16.3.1
