Dataset Slices

​

Creating Your First Simple Slice

Imagine you want to monitor model performance on samples containing the keyword “star wars.” To do so, you can simply type “star wars” into the search panel and save the resulting data as a new custom Slice (see Figure below).

When creating a new slice you are presented a pop up that allows you to give a custom name to your slice and displays slice level details: 1) Slice project scope, 2) Slice Recipe (filter rules to create the slice). Your newly created slice will be available across all training runs within the selected project.

​

Complex Slices

You can create a custom slice in many different ways e.g. using similarity search, using subsets etc. Moreover, you can create complex slices based on multiple filtering criteria. For example, the figure below walks through creating a slice by first using similarity search and then filtering for samples that contain the keyword “worst.”

The final “Slice Recipe” is as follows: