Skip to main content
Fine-tuning adapts an open base model to your own data. Valkyrie runs the training for you on an on-demand GPU using a managed training pipeline, then tests the result against the base model so you can see how much better it got.

​
Two ways to start

Chat with the AI assistant

Attach a spreadsheet or text and say what you need. The assistant builds the training data, picks the model, prices the run and prepares it for you to confirm.

Bring your own examples

Pick a dataset, a base model and a recipe yourself, step by step.

​
Before you start

  • Your data: a dataset already in training format, a table or text for the AI assistant to turn into one, or an example set.
  • Enough wallet balance to cover the training run. You see the time, the cost and your balance after the run before anything starts. For a model under Ready to fine-tune, the wallet must hold the estimate’s high end plus a 20% buffer, see Before a fine-tune starts.

​
Models and versions

Every fine-tune is a version (v1, v2, and so on) of a model. When you start one, choose New model or New version of a model. A version can use different examples or another base model, so train again with New version to improve a model and every result stays side by side on the Models screen.

​
Create a fine-tuning job

1

Open New fine-tune

Click New fine-tune in the top bar. You can also start from Home with I have my examples (pick a dataset, base model, and recipe, then click Start fine-tune), or from Fine-tune with these on a dataset’s page.
2

Pick your examples

Under Your examples, choose a file from My datasets, or Upload new. Test it on holds back 10% or 20% of them to test the result against the base model. Under Checks and targets, pick the checks the result must pass, such as Valid JSON or Label accuracy, and a target for each.
3

Choose a base model

Pick one of the models under Ready to fine-tune (Qwen2.5 7B Instruct is a good default), or enter Any Hugging Face model by its owner/name id. See Model names. Each model’s card shows the typical time and cost for 1,000 rows.
4

Choose a recipe

Quick check (1 pass) tells you fast whether your data teaches anything, Standard uses settings tuned for the base model, and Thorough (5 passes) suits larger datasets. Advanced settings has the full training configuration. Tick Interruptible (cheaper bid box) to train on cheaper GPU time that can be taken back, see Interruptible capacity.
5

Start

In an organization workspace, choose the Owner: you, a team, or the workspace. See Who owns new work. Check the Summary, with its time and cost Estimate and your Balance after, and click Start fine-tune. Training is billed per minute and you can cancel any time. Valkyrie schedules the job, provisions a GPU, and runs the training. If a preferred GPU type isn’t available, provisioning retries with fallbacks.
See New fine-tune for a screen-by-screen tour.

​
Track progress

A job moves through a series of states as it runs:
Initializing covers GPU start-up, model download, and dataset prep; Running is active training; Finalizing is the job saving and uploading its outputs after the last step. After you start, you land on the new version’s page under Models, where you can watch the stages and live progress or cancel the run. The Pipeline view on Models shows every model’s current step at a glance. Logs are also available through the MCP logging tools.

​
Outputs

When a job completes, its outputs (the trained weights or adapters) are stored for you. From the completed version on Models you can:
  • See its Test results: how well your version and the base model answer the held-back examples, and the gain in points, plus the Checks you picked and how many targets were met.
  • See the Training record and Download the outputs with temporary signed URLs.
  • Deploy the version, see Deploy a fine-tuned version.

​
If a job fails

Fine-tuning runs on ephemeral cloud GPUs, so transient infrastructure issues can occur. A failed version explains what went wrong in plain language. Click Retry to run it again without recreating it, or Change settings and try again to adjust it first. Expand Technical detail if you need the raw error. Retrying usually reschedules the job onto fresh hardware. A job started on interruptible capacity resumes from its last checkpoint on a new machine when it’s outbid, up to 3 times. After that it fails, and Retry continues from the last checkpoint. Starting a fine-tune needs a role that can spend (Owner, Admin, or Builder) and room under the workspace’s or team’s spend limit.
Fine-tuning has a maximum run time. Very large datasets or long training configurations should be sized with that limit in mind.

​
Archive old versions

Keep your list tidy by archiving versions or whole models you no longer need. On Models, select the model, click Archive, and choose the selected version or the whole model. Nothing is deleted: outputs, billing, and usage history stay, and Show archived lets you Unarchive them later. A version that is still training, or that has a running endpoint, must be stopped first. See Archive a model or version.

​
Deploy what you trained

Deploy a fine-tuned version

Turn a completed version into a live, OpenAI-compatible endpoint.