Relabeling - CoreWeave Docs

Relabeling asks a configured model to produce a new answer for every dataset entry. The result is a separate output set. Original responses remain available, and multiple relabel runs can coexist.

When relabeling helps

Relabel when the production model output is not the behavior you want the student to learn, for example:

A stronger model can correct weak or inconsistent answers.
The instructions or structured-output requirements have changed.
Production data contains acceptable inputs but noisy labels.
You want to compare two teachers before training.

Skip relabeling when the current outputs are already the desired targets. It adds inference cost and can transfer the teacher’s mistakes or style unnecessarily.

Choose a relabeling strategy

The following sections describe the two relabeling strategies.

Regenerate outputs

The model receives each request exactly as the original model did but doesn’t receive the existing output. It then produces a fresh answer. Use this strategy when you want different model behavior, such as a stronger teacher or a model that follows updated instructions.

Create revised outputs

The model receives the original answer and your relabeling instructions as context. It then produces a corrected version. Use this strategy when the original outputs are mostly correct and you want targeted fixes that preserve their intent, style, structure, and tool-use decisions. Revised runs accept two additional settings:

Relabeling instructions: How the model should treat the original output. By default, the model treats the original as correct unless it contains a clear, material error. When the original contains an error, the model makes the smallest correction that preserves the original’s intent, style, structure, and tool-use decision. Instructions can contain up to 20,000 characters.
Reference conversations: Complete conversations selected at random from the training split that provide examples of the task. Each request includes these conversations. The default is 10, and the maximum is 50. To omit them, use 0. Studio shows the estimated additional input tokens per request. The examples form a stable prefix that providers can cache.

Run a relabel

To run a relabel, complete the following steps:

On the project’s Datasets page, select a ready dataset.
Select Relabel. Under Relabel with…, choose a model.
Choose a Relabeling strategy.
For Create revised outputs, review the reference conversation count and the relabeling instructions.
Optional: If you published the dataset revision to W&B Weave, select Publish outputs to Weave to publish the finished outputs beside it.
To start the relabel run, select Relabel dataset.

The background job processes entries concurrently and reports its state as queued, running, completed, failed, or stale. After the relabel run completes, review the results:

Inspect original and relabeled outputs side by side.
Filter unchanged or errored rows.
Select the relabel run as the training output when creating a fine-tune.
Use it as the primary reference in a comparative evaluation.

Deleting a relabel run removes only its generated output set. It does not delete the dataset, original outputs, source traces, or provider.

API: Relabel a dataset (POST /tasks/{alias}/datasets/{datasetId}/relabels)

Replace the dataset ID and choose any registered provider model. To regenerate outputs:

curl --request POST \
  --url "https://distillation.training.wandb.ai/v1/tasks/ticket-classifier/datasets/$DATASET_ID/relabels" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Wandb-Entity: your-team" \
  --header "Content-Type: application/json" \
  --data '{"model_ref": "openai/gpt-5.6-sol"}'

To create revised outputs, set use_original_output to true. You can omit instructions and reference_conversation_count to use the defaults described earlier:

curl --request POST \
  --url "https://distillation.training.wandb.ai/v1/tasks/ticket-classifier/datasets/$DATASET_ID/relabels" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Wandb-Entity: your-team" \
  --header "Content-Type: application/json" \
  --data '{
    "model_ref": "openai/gpt-5.6-sol",
    "use_original_output": true,
    "reference_conversation_count": 10,
    "instructions": "Treat the original output as correct unless it contains a clear, material error. If needed, make the smallest correction while preserving its intent, style, structure, and tool-use decision."
  }'

To publish the finished outputs to Weave beside the dataset’s published revision, add "publish_to_weave": true to either request.The response contains the relabel run ID and status. See Start a relabel run.

​When relabeling helps

​Choose a relabeling strategy

​Regenerate outputs

​Create revised outputs

​Run a relabel

When relabeling helps

Choose a relabeling strategy

Regenerate outputs

Create revised outputs

Run a relabel