Datasets

The Datasets section is where you upload and manage the data your training jobs run on. It mirrors client.datasets in the SDK.

Upload a dataset

Click Upload dataset and select a local .jsonl file. Fleet uploads it to object storage and lists it with its name, size, and upload time.

Datasets are content-hashed, so uploading the same file twice doesn't duplicate storage.

Dataset format

Training expects newline-delimited JSON (.jsonl), one example per line. The exact schema depends on your training method; for chat-style fine-tuning each line is typically a conversation with a messages array. See Training for method-specific details.

Using a dataset

When you start a training job from the Models section, your uploaded datasets appear in the dataset picker. Selecting one also unlocks the live cost and time estimate for the run, since Fleet uses the dataset's row count to project how long training will take.

Manage datasets

Each row offers a delete action. Datasets are billed as stored data while they exist; deleting one frees its storage immediately. See Storage for how storage is metered.