I feel like the data should have been generated by a much less predictable policy.
It often feels like the model is ignoring my inputs and just doing what it would expect the bot to do (which is unsurprising if the model could predict what would happen next during training without paying attention to the inputs)