Getting different accuracy each time (randomness)

Rasa is awesome, but I’m running into some pretty weird issues. Every time I train Rasa Core, the model trained (via Keras) outputs a different final accuracy. I searched all over and couldn’t find an answer, so please forgive me if this question has been answered.

It’s troublesome because the same stories.md file will sometimes work and sometimes not. I notice that it reports a sample size that’s a lot larger, so I can only assume that Rasa does some sort of preprocessing. However, the indeterministic nature of the training makes it very difficult to send our training data to our various build servers and expect to get the same results (e.g. QA vs prod).

Is there some way to stop whatever is causing the randomness?

To others facing a similar issue, I was able to have the different build servers converge more or less to the same dialogue paths by increasing the number of epochs that the RNN is trained for in policies.yml.