A training failure where generated sequences become so long they get cut off, biasing the training data toward incomplete examples.