How many stories are needed for a model?

Rasa Open Source
1

The core model that predicts next action presumably trains only on stories. So if I have only 10 stories how can that be enough to train any kind of model? Is there a minimum number of stories required?

I spent some time creating a few short stories. In some cases I followed a story exactly yet the prediction was not the next step in the story but a completely unrelated action. I fixed by creating some new stories using rasa interactive. These are much more complex and quite hard to follow but seem to work! Is the right approach to just fix in rasa interactive and try not to overthink the logic?

1 Like
2

There isn’t really a specific number.

What I usually do is write rules which are just a direct intent-to-action link, as well as rules for forms.

For more complicated stuff, I start writing minimal stories as a few relevant intent-action chain, as well as a bit more complicated stories for forms.

Finally, I use interactive learning to build more stories based on read conversations until predictions gets good.

Furthermore, don’t forget you can tweak your policies to increase accuracy - I suggest using Tensorboard with Rasa to compare pipelines and policies

2 Likes