Datasets Quick Start - CoreWeave Docs

This quick start shows you how to build a dataset from the traces a project has already recorded. The result is a set of training and test rows for relabeling, fine-tuning, and evaluations.

Before you begin, send representative traffic through the project. Studio builds datasets from the project’s own traces. To build a dataset from W&B Weave calls logged outside the project, use the API source described in Datasets.

Open the project

Select the project you want to improve, open Datasets, and select New Dataset.

Name the dataset

In Dataset name, keep the generated name or enter your own. The dataset is a self-contained snapshot of traces logged to this project in W&B Weave.

Choose the source data

Under Source data, configure the following fields:

Project version: The version whose prompt and routing regime produced the traces you want. Studio selects the latest version by default.
Served by: The model whose responses become the training targets. To keep the training data teacher-only, select the teacher model. All models mixes the outputs of every model that served the version, which includes any fine-tuned student.
After (UTC) and Before (UTC): The fixed time window to select traces from. Studio prefills the last 14 days as fixed timestamps.
Maximum traces: The random sample size after filtering. The default is 20,000 and the maximum is 100,000. Each selected trace becomes one dataset entry.

Set the train and test split

To choose how many rows to reserve for evaluations, use the Train / test split slider. The default is 90% training and 10% test. Each complete trace stays in one split, so evaluations don’t include near-duplicates from training.

Create and inspect

Select Create Dataset and wait for the status to become Ready. Review several training and test rows, check message and output fidelity, and delete unusable entries before training.

The dataset is now ready for relabeling, fine-tuning, and evaluations.

API: Create a dataset (POST /tasks/{alias}/datasets)

Use fixed timestamps so rerunning your workflow describes the same source window. The filters object corresponds to the Source data fields in Studio: task_version is the project version, and resolved_provider with resolved_model is the Served by model. To include all models, omit those two keys:

curl --request POST \
  --url "https://distillation.training.wandb.ai/v1/tasks/ticket-classifier/datasets" \
  --header "Authorization: Bearer $WANDB_API_KEY" \
  --header "Wandb-Entity: your-team" \
  --header "Content-Type: application/json" \
  --data '{
    "name": "recent-production",
    "params": {
      "query": {
        "source": "task",
        "after": "2026-08-01T00:00:00.000Z",
        "before": "2026-09-01T00:00:00.000Z",
        "filters": {
          "task_version": 1,
          "resolved_provider": "openai",
          "resolved_model": "gpt-5.6-sol"
        }
      },
      "sampling": {"strategy": "random", "limit": 20000},
      "split": {"val": 0.1, "by": "trace"}
    }
  }'

The response contains the dataset ID. Poll the dataset resource until status is ready. For details, see Create a dataset.

Last modified on September 2, 2026