Training loss continues to fall while validation loss rises for several epochs. What does this pattern most strongly suggest?

1. The tokenizer must have exactly one token per sentence.
2. The model is increasingly overfitting the training data.
3. The model has achieved perfect generalization.
4. The validation set has become part of the training set by definition.