Also, while it's clearly got a lot of training on some use cases, others that probably weren't in the training set have worse good decision rates than a random number generator. If you can rebuild the architecture, you can train it on your use case.