Leaving OpenAI fine-tuning?
OpenAI’s deprecations page lists 6 January 2027 as the date OpenAI stops self-serve fine-tuning job creation. Read OpenAI’s notice.
46 days left.
The check runs in this browser. Your file does not leave this device: nothing is uploaded, and only the counts you choose to attach to a request are sent.
What a team keeps
- The training data, in your own project, frozen as versions after review. Your OpenAI chat files train as they are, with no conversion.
- The trained adapter, written on your own machine by the Valcairn runner, and an export you can load into Ollama there.
- The record: which data version, which settings, and the held-out loss before and after training.
What the check looks at
The same rules the training runner uses: lines it would train on, lines it would skip and why, how many distinct prompts there are (a job needs at least five), and roughly how many examples it would hold out. Tool calls and image parts are not read today; those lines are listed as skipped. Files up to 2 GB each, 20 GB per training job. The training file format.
Early access: requests are reviewed by hand, and dataset approval is a staff step. A request is not a booking or a purchase.