Datasets
Datasets will be where the data behind fine-tuning and evaluation lives — uploaded once, versioned, and reusable across jobs.
How we expect it to work
Section titled “How we expect it to work”- Upload a file of examples, most likely JSONL with one example per line.
- Validate — we check the shape and report bad rows before you spend GPU time on them.
- Split into training and evaluation sets.
- Version, so a training run always records exactly which data produced it.
What we expect to hold true
Section titled “What we expect to hold true”- Datasets are scoped to a workspace and visible only to its members.
- Your data is used for the jobs you start, and nothing else.
- Deleting a dataset does not retroactively untrain a model already built from it — you would retrain or delete that model too.
Reproducibility is the point of versioning here: months later, you should still be able to answer exactly what a given model was trained on.