You can add datapoints either from file or by manually adding datapoints one-by-one.

​
1. Export from a span

You can export a span as a datapoint from both Traces and Evaluation traces. To do so, select a span and click “Export to dataset” in the top right corner.

You will be prompted to modify or keep only the necessary fields. Select the dataset you want to export to and click “Export”.

This will fill the data field of the datapoint with the selected attributes of the span.

​
2. File upload

You can upload datapoints from a structured file with datapoints.

To do that, click “Add from source” at the top of the Datasets page. Then, select in the tab whether you want to upload in a structured or unstructured way.

​
File format

Supported file formats are: .csv, .json, .jsonl. We infer the format based on the file extension.

  • csv - header is required, default separator and minimal quoting are assumed. If a row has an empty value or less values than headers, missing values will be filled with empty strings.
  • json - the file must contain one array of datapoints.
  • jsonlines - every line must contain one datapoint.

For each datapoint, we first construct the key-value object, and then parse it according to the following rules:

  1. If the only key is "data", and the corresponding value is a JSON object, place that object in datapoint’s "data".
  2. If the keys are "data" and "target", and both corresponding values are JSON objects, populate the datapoint’s "data" and "target" accordingly.
  3. In ALL other cases, all keys and values will go inside "data".

This is needed to ensure that at the top level "data" and "target" are both JSON objects.

If there is an error parsing the file, no datapoints will be added. If a single value in the file does not conform to format, it will be silently ignored.

​
3. Add manually

Click “Add datapoint” at the top of the Datasets page and new empty row will be added. Then, you can fill out the content of “data” and “target” for newly created datapoint. Make sure to keep both data and target as JSON objects.

Label spansIntroduction