Fine-tuning Quick Start - CoreWeave Docs

This quick start shows you how to train a task-specific model from a ready dataset in Studio and use the deployed artifact for evaluation and routing.

Before you begin, create a dataset with training entries. If you generated better targets with relabeling, wait for that relabel run to complete.

Open Tuned Models

From the project, open Tuned Models and select Fine Tune.

Choose the training data

Select a ready dataset, then choose the original outputs or a completed relabel run as the targets the model should learn.

Select a base model

Choose one of the supported training models. When compute is available, train more than one candidate instead of assuming the largest model will perform best.

Keep or customize the defaults

The fine-tune name is optional. For the first run, leave the training configuration automatic. To configure the run, expand the customization options. You can cap the training examples, set the batch size and epochs, and specify a peak learning rate or a full learning-rate schedule. You can also choose which assistant turns to train on or pass experimental W&B Serverless Training options. See Training parameters for details.

Start training

Select Start training. The model moves through Queued and Training, and Studio reports progress from 0 to 100%.

Use the deployed artifact

When the status becomes Deployed, the exact W&B artifact appears in Studio and becomes available for evaluations and routing through wandb-inference.

Next, evaluate the fine-tune on held-out validation rows before sending production traffic to it.

API: Start training (POST /tasks/{alias}/finetunes)

Omit model_config to use the automatic training configuration:

curl --request POST \
  --url "https://distillation.training.wandb.ai/v1/tasks/ticket-classifier/finetunes" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Wandb-Entity: your-team" \
  --header "Content-Type: application/json" \
  --data "{
    \"dataset_id\": \"$DATASET_ID\",
    \"model_name\": \"ticket-classifier-qwen\",
    \"base_model\": \"Qwen/Qwen3.6-27B\"
  }"

Poll the returned fine-tune until it is deployed or failed. See Create a fine-tune.

Last modified on September 2, 2026