It’s 10+ years old, freely licensed and available in multiple languages. So it should very likely be in the training data. They could even train it with the sources and its layers