Datasets and Exercises - where to find them

Initial Build Exercises#

The starting exercises need only a blank workspace connected to the ABCD dataset.

  • Click "Create Workspace">"Create a Blank Workspace"
  • Give it a descriptive name of your choice
  • Go to "Data sources" near the middle of your screen
  • Select "Use one of our public data sets" you'll see we (arbitarily) split ABCD is split into two fictious months
    • "ABCD 👚 (May 2022)"
    • "ABCD 👚 (June 2022)"
  • Each has about 5000 conversations in, most exercises assume you're connected to the May data
  • The June data can then be used in continuous improvement scenarios against a model trained on May
  • You can connect both if you like
  • You won't see the dataset in "Data Management" as it's centrally hosted - this way it doesn't count against your data point allowance
  • When you create labelled phrases for exercises or add new utterances these will count to datapoints
  • You can always remove the workspace when finished with the extercise.

Now head over to your first example to see how to investigate the data.

There are no particular challenges or exercises set for the first exercises, it's just about exploring the data and learning the tool.

Exercises requiring a trained workspace#

If you want to start with a pre-existing model we provide workspaces built for a number of different scenarios. These are all based on the sort of model you might create on ABCD growing from around 30 intents upwards.

To select an exercise:

  • Click "Create Workspace">"From a demo Workspace"
  • Give it a descriptive name of your choice
  • If you go to intents you should see some trained data
  • Attach the unlabelled ABCD data set for the exercise to it using "Data Sources"

These models all come with a particular challenge based around a flaw in them. The description should tell you what that is.

You can either head over to the improve a model exercises which will take you through the techniques it aims to cover and outline what the exercise invovles with some videos. Or if you are familiar with the tool, just jump in and try and solve them.

To get the answer, there is a click to reveal section at the bottom of each exercise.

In general each exercise builds on the last, fixing the issue and moving the model forward in other ways. Each stage represents a key stage we find customers reach in NLU Design and shows the evolution of a model based on ABCD.