MANAGE DATASETS

After you create a dataset, review its rows before you train on it, and keep track of its versions as you change it.

​

REVIEW ROWS IN DATA EXPLORER

Data Explorer lists every row of the dataset with its content and metadata. Click Data Explorer on the dataset page, or select it from the dataset’s menu on the Datasets page. Click a row to read the whole conversation, and click View metadata to see its metadata. Before you use a dataset, read a sample of rows and check that:

• Training rows end with an assistant turn. Training learns from the assistant’s responses, so a row that ends on a user turn teaches nothing.
• Labels match the responses. For a classification dataset, the label in metadata should match the assistant’s answer. If they disagree, the evaluation scores measure the labeling errors instead of the model.
• Generated rows vary. In synthesized data, scan the assistant turns for repeated openings or structure. A model trained on repetitive data learns one template.
• Tool calls are present. In an agent dataset, an assistant turn that calls a tool shows its calls, and its text can be empty.

Data Explorer is read-only. To change rows, select Download… from the dataset’s menu to export the latest version as JSONL. Edit the file, and upload it as a new dataset.

​

VERSIONS

Oumi creates a new version of a dataset when you delete failed rows, restore a version, or add generated rows to it. The Versions tab lists every version, marks the Current Version, and shows when each one was created.

• Restore a version: Open the menu of an earlier version and select Restore Version. Oumi copies that version into a new version and keeps the current one, so you can undo the restore the same way.
• Delete a version: Open the version’s menu and select Delete. If you delete the current version, the previous version becomes current. You can’t delete a dataset’s only version.

To view an earlier version without restoring it, choose it in the Version selector on the Overview tab or in Data Explorer.

​

DELETE A DATASET

To delete a dataset, open its menu and select Delete. Deleting removes the data and every version, and it can’t be undone. Before you delete a dataset, check the Used in card on its Overview tab for the trainings and evaluations that used it.