I would assume this is simply not enough data. You also have access to all public domain books and they usually make up a small fraction of the data used to tra

I would assume this is simply not enough data. You also have access to all public domain books and they usually make up a small fraction of the data used to train a model from scratch. For fine tuning a model, having a unique source of high quality data is probably valuable even if small.