Datasets
Datasets are the curated collections of examples you use to fine-tune models in Rational AI. Each dataset stores its entries in a standard training format, keeps a full version history as you add more data, and can be downloaded or reused across fine-tuning jobs. Use this section to build, organise and maintain the training data that powers your custom models.
Available on any plan
Anyone with can edit access can create and manage datasets
Datasets are used to fine-tune models, which is only available with Pro Plans
Supported formats
When you create a dataset you choose the format its entries follow. The structure of any file you upload must match the selected format:
| Format | Description |
|---|---|
| Alpaca | Instruction-style entries, each pairing an instruction (and optional input) with the expected response. Best for single-turn, task-oriented fine-tuning. |
| ShareGPT | Conversation-style entries, each containing a list of turns labelled by role (e.g. human, gpt). Best for multi-turn, chat-oriented fine-tuning. |
Create a dataset
Click New in the top-right corner and choose how you want to build the dataset:
- From file – Upload an existing JSON file of training entries.
- From events – Generate a dataset from conversational events already captured in the platform.
Create from file
- Choose From file
- Enter a Name and an optional Description
- Pick the Format (Alpaca or ShareGPT) that matches your file
- Drag a file into the upload area, or click Click or drag here your dataset to select one from your computer
- Click Save
The uploaded file must follow the structure of the selected format, otherwise the import is rejected.
Create from events
Choose From events to assemble a dataset from conversations and interactions that have already taken place in Rational AI. This is a convenient way to turn real usage into training data without exporting and reformatting it by hand.
Work with a dataset
Click a dataset row to open its detail view. The page shows the dataset name, its description and a paginated table of entries under the content column. Each row is one training entry stored as JSON; click the arrow at the start of a row to expand it and inspect the full entry.
Manage versions
Every dataset keeps a version history. The version selector in the top-right corner (for example, v1) lets you switch between versions. A new version is created each time you append entries, so you can always trace how the dataset evolved and roll back to an earlier state if needed.
Append new entries
Use Append new entries (from the row menu or the detail-view menu) to add more examples to an existing dataset. Appending entries produces a new version while preserving the previous one.
Manage datasets
Open the options menu (the three vertical dots) on any dataset row, or the menu in the detail view, to access the following actions:
| Action | Description |
|---|---|
| Append new entries | Add more entries to the dataset, creating a new version. |
| Rename | Change the dataset's name. |
| Download | Export the dataset to a JSON file in its current format. |
| Delete | Permanently remove the dataset. |
Browse datasets
The Datasets list shows every dataset with columns for Name, Format, Entries (the number of training examples), Created at, Updated at and Current version.
- Search – Use the search box above the list to filter datasets by name.
- Sort – Click the arrows next to any column header to sort the list by that column.