> ## Documentation Index
> Fetch the complete documentation index at: https://docs.coreweave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Fine-tuning

> Train supported base models from original or relabeled dataset outputs.

Fine-tuning creates a task-specific model through W\&B Serverless Training and records the resulting hosted artifact in the UI.

> [card] [Train your first model](https://docs.coreweave.com/model-distillation/studio/fine-tuning-quickstart.md)
>
> Choose a dataset, base model, and training configuration in the UI.

## Supported base models

For available models and selection guidance, see [Supported training models](https://docs.coreweave.com/model-distillation/supported-models.md). The UI can display the base model for an existing fine-tune even when the model is no longer available for new training runs.

## Create a fine-tune

Choose the following:

* A ready dataset with training rows.
* Original outputs or a completed relabel run.
* A base model.
* An optional fine-tune name.
* Optional training parameters.

Names are normalized to lowercase hyphenated identifiers. When you omit a name, the UI generates a readable one.

## Training parameters

Open the advanced configuration only when you need control over:

| Field | Meaning |
| - | - |
| Maximum examples | Deterministic cap on training rows. Validation is not capped. |
| Batch size | `auto`, which resolves to 2, or an integer from 1 through 32. The UI offers Auto, 1, 2, and 4 for most models, and up to 32 for the Nemotron model. |
| Epochs | Number of complete passes over the training set. Defaults to 1. |
| Peak learning rate | Positive scalar used to build the default schedule |
| Learning-rate schedule | Explicit non-negative value for every optimizer step |
| Assistant turns | Train on all assistant turns or only the last turn |
| Experimental configuration | Serverless Training options not yet modeled as first-class fields |

With a scalar peak learning rate, Model Distillation creates the same schedule used by the ART SFT helper: 10% linear warmup followed by linear decay to zero. Supplying a full list preserves it exactly.

## Status and artifacts

Fine-tunes move through **Queued**, **Training**, **Deployed**, or **Failed**. Training progress is reported from 0 to 100%. A deployed model includes an exact W\&B artifact reference and is immediately selectable through `wandb-inference` for evaluations and routing.

A failed fine-tune shows one of two actions next to its status, **Retry** or **Create new model**, depending on where the failure happened.

If the failure happened in the Model Distillation pipeline itself, such as during preparation, while monitoring the training job, or while recording the result, select **Retry**. What the retry does depends on how far the fine-tune progressed before the failure:

* If no training model was created yet, the retry restarts preparation.
* If a confirmed training job exists, the retry resumes it with fresh credentials, keeps the completed epochs, and trains only the remaining ones.
* If a training model exists but no confirmed submission can be recovered, the retry returns `409 Conflict`, so create a new fine-tune instead.

If the training service itself reported the failure, the job can't be resumed, and the UI shows **Create new model** instead. This action opens the fine-tune dialog prefilled with the failed fine-tune's dataset, output source, base model, and training settings, together with the failure reason. You can adjust the settings before you train again.

In the API, a training-service failure sets the fine-tune's `training_failure_type` to `RemoteTrainingFailed`, and a retry returns `409 Conflict` with type `remote_training_failed`. A retry also returns `409 Conflict` in any of the following cases:

* The fine-tune isn't in the failed state.
* A retry that restarts preparation detects that the dataset changed after you created the fine-tune.
* A training model exists but no confirmed submission can be recovered.

## Delete a fine-tune

Deleting a fine-tune removes the training model, its W\&B artifacts, its generated outputs, and the UI record. A training model shared by several fine-tunes is removed when you delete its last owning fine-tune.

Model Distillation removes the fine-tune from evaluation comparison sets and deletes evaluations that use it as ground truth. For a head-to-head evaluation that uses it as the primary participant, Model Distillation promotes the oldest remaining participant and marks the evaluation stale. An active evaluation blocks deletion.

To delete a fine-tune in the UI:

1. Open the fine-tune's page.
2. Under **Danger Zone**, select **Delete Fine Tune**.
3. In the dialog, review the affected evaluations and type the model name to confirm.
4. Select **Delete**.

Deletion runs in the background. If it fails before any destructive work begins, the model becomes available again. If it fails afterward, the model stays locked until a retry succeeds.

To retry, confirm the deletion again. Retries stay within the scope that the original request recorded.

> [accordion] API: Delete a fine-tune (DELETE /projects/\{alias}/finetunes/\{finetuneId})
>
> Deletion takes one request, and the server determines the affected evaluations and blockers itself:
>
> ```bash theme={"system"}
> curl --request DELETE \
>   --url "https://forge.coreweave.com/api/distillation/v1/projects/ticket-classifier/finetunes/$FINETUNE_ID" \
>   --header "Authorization: Bearer $WANDB_API_KEY" \
>   --header "Forge-Entity: your-team"
> ```
>
> The request returns `202 Accepted` with the deletion operation and its status. If you repeat the request while the operation is queued, running, or completed, the request returns the same operation. A repeated request also retries failed cleanup within the scope that the original request recorded. When the fine-tune is already gone and no operation exists, the request returns `204 No Content`, and `GET` on the fine-tune returns `404 Not Found`.
>
> To preview the affected evaluations and blockers before you delete a fine-tune, call `GET /projects/{alias}/finetunes/{finetuneId}/deletion-plan`. The preview is informational and also reports the status of a running deletion. See [Delete a fine-tune](https://docs.coreweave.com/model-distillation/reference/management/fine-tunes/delete-a-fine-tune.md) and [Preview fine-tune deletion](https://docs.coreweave.com/model-distillation/reference/management/fine-tunes/preview-fine-tune-deletion.md).

> [accordion] API: List training jobs and artifacts (GET /projects/\{alias}/finetunes)
>
> To check progress, terminal status, and the artifact reference used in evaluations and routing, use this request:
>
> ```bash theme={"system"}
> curl --url "https://forge.coreweave.com/api/distillation/v1/projects/ticket-classifier/finetunes" \
>   --header "Authorization: Bearer $WANDB_API_KEY" \
>   --header "Forge-Entity: your-team"
> ```
>
> To start a job, use the request in [Fine-tuning Quick Start](https://docs.coreweave.com/model-distillation/studio/fine-tuning-quickstart.md). For this operation, see [List fine-tunes](https://docs.coreweave.com/model-distillation/reference/management/fine-tunes/list-fine-tunes.md).
