Make rasa train reproducible

Every time I train the model using the rasa train command, I get the model with slightly different behavior. Was wondering if there is any way to set a random seed during training so that every time we train on the same nlu files, we get the exact same model.

1 Like

Looks like we can set an argument under DIETClassifier in config.yml - random_seed

Hi, yes you can set random seeds for some parts of the pipeline. See this thread here for an example config: Trained model is very inconsistent You can also set it for TEDPolicy if you are using that in your config.

that’s nice, will try this out. thanks

Yup :slight_smile: You can also set the seed for ResponseSelector and TEDPolicy like you can see here.

1 Like