Datasets Quick Start - CoreWeave Docs

Before you begin, send representative traffic through a task or identify an existing W&B project containing compatible Weave calls.

Open the task

Select the task you want to improve, open Datasets, and choose Create dataset.

Choose the source

Use Task traffic for calls captured by the Model Distillation proxy. Select Existing W&B project only for compatible calls logged outside the task.

Define a fixed snapshot

Choose After and Before dates, the task version, and the provider/model that produced the behavior you want to learn. Set a maximum only when you want to sample a larger result set.

Reserve validation data

Start with 20% validation data. Model Distillation keeps related scenarios, traces, and conversations in the same split so the evaluation does not see near-duplicates from training.

Create and inspect

Start the build and wait for Ready. Review several train and validation rows, check message and output fidelity, and delete clearly 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:

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}
      },
      "sampling": {"strategy": "random", "limit": 20000},
      "split": {"val": 0.2, "by": "trace"}
    }
  }'

The response contains the dataset ID. Poll its returned resource until status is ready; see Create a dataset.

Last modified on August 25, 2026