Datasets are saved sets of content you can test and evaluate policies against. The Manage Datasets page is where you upload, edit, label, and export them.

Upload a dataset

Upload a CSV file with your content. You can either:

• use columns named input and expected_output, which are detected automatically, or
• select which columns to map to input and expected output.

All other columns are stored as metadata.

Datasets can include images, videos, or sound files as well as text: upload a CSV that references media files, and they render as previews in the dataset viewer.

When creating a dataset, there are four Dataset Type options to choose from: Training, Golden, Moderator Decisions, and Custom. This doesn’t materially change anything, it just affects how the dataset is organized in the main Dataset table. You can sort alphabetically in the “Type” column to see all the same-type datasets grouped together.

You can also create a dataset via API if you’d prefer to ingest data rather than upload a CSV.

Translate Rows

Look for the small translation icon next to text rows — click on it to translate the row to English.

Edit a dataset

Click into the dataset to change it directly. You can change the expected output (the decision or label), and add or remove rows.

Clicking the … at the top of a dataset table will show you Utilities, which include renaming the dataset, editing the schema, exporting the full dataset, sharing, or deleting.

Relabel many rows at once

Two options when a single-row edit is too slow: