Skip to main content

Training Process

Open the model's details page (/custom-models/<id>) - uploading datasets and running training both happen here.

Datasets Section​

Below the status banner is the Datasets section: a title with an info tooltip about dataset types, a caption, and a New dataset button on the right.

Datasets section header Section header - title, tooltip, New dataset button.

Upload a Dataset​

Click New dataset. The dialog shows the model name and a Type selector with five options:

  • Auto-detect (default) - the server picks the type per file from its extension.
  • TRANSCRIPT - restricts the picker to transcript formats.
  • TEST - same transcript formats, intended for validation data.
  • AUDIO - restricts the picker to audio formats.
  • MANIFEST - restricts the picker to .jsonl manifests.

Drag files into the dashed area or click Choose files. Each batch can have 1-10 files; you can keep adding more batches afterwards. The picker's accept attribute is set from your type choice (or the union of all formats on Auto-detect), so files outside the supported extensions can't be picked.

New dataset dialog with files selected Type selector, drop area with files queued, Create button.

Click Create. The button shows Uploading... while the request runs; the type selector and Cancel are locked during the upload.

The upload goes in a single multipart PUT request - there is no chunking or resumable client. If the connection drops, retry the batch.

When the request finishes you get one toast:

  • 1 file → Dataset uploaded successfully.
  • N files → N files uploaded successfully.
  • Failure → the error message returned by the server.

Supported Formats​

  • Audio: .wav, .mp3, .m4a, .flac
  • Transcripts (TRANSCRIPT and TEST types): .vtt, .srt, .txt
  • Manifest: .jsonl

Datasets Table​

Each uploaded file appears as a row with Name, Type, Duration, URL, Start time, End time. The Name cell copies the file name to your clipboard on click. The row menu has a single Delete action - confirm and you'll see a File deleted successfully toast.

Datasets table with uploaded rows Uploaded dataset files.

Don't have timestamps in your transcripts? Use Force Alignment from the dashboard's Create button to add timestamps before uploading. The Need help preparing your data? link on the status banner takes you back to the dashboard.

Start Training​

Click Train the language in the top right of the details page.

The button is only visible when the training status is Not running (1) or Failed (5). It is hidden for Ready to run (2), Running (3) and Success (4).

A confirmation dialog opens with the model's name and a warning that training uses your credits and the model can't be modified while it runs. Click Train to confirm.

Train confirmation dialog Confirmation dialog before training starts.

What Happens Next​

  1. The panel calls validate (POST /api/v3/custom_models/<id>/validate). If validation fails, the server's error message is shown as a toast and training does not start.
  2. If validation passes, the panel calls train. On success you get a Training started successfully toast and the status updates on the next refresh (within 20 s).

Recognised Error Codes​

The panel re-labels these codes from the server:

  • validation_failed - datasets don't meet the server-side training requirements.
  • invalid_audio_format - at least one audio file uses an unsupported format.
  • invalid_transcript_format - at least one transcript file uses an unsupported format.
  • no-claims-available - your plan's training quota is exhausted.

Anything else is shown verbatim from the server's message / error_description.

Training Statuses​

CodeLabelColourWhat it means
1Not runninggreyNo training has started, or you can start a new one.
2Ready to runamberTraining is queued and will begin shortly.
3Runningsky blueTraining in progress. May take several hours.
4SuccessgreenTrained - the model is usable on new transcriptions.
5FailedredSomething went wrong. Review the datasets and try again.

The status banner on the detail page shows a help message tailored to the current status. It auto-expands when the status is Success so the message that names your model is visible without a click.

Training status badges The five status badges.

Next Steps​