Work with datasets

A Dataset is an App-scoped collection of input and expected-output pairs. You can build one from production Logs, import existing examples, or add items manually. The same Dataset can then supply test cases to Evaluate and training examples to Fine-tune.

Before you begin

You need a Workspace and an App. To create a Dataset from existing Logs, your App must also have generated data. To import a file, prepare a .csv or .jsonl file whose rows contain valid input and output data.

For instructions on logging, filtering, and collecting feedback, see Manage Logs. This guide focuses on the Dataset lifecycle in the product.

Create a Dataset from Logs

Use this flow when you want a reproducible collection from production or test generations.

  1. Open your App and select Logs.
  2. Filter the table, or select individual rows.
  3. Select Save Dataset.
  4. Enter a Dataset Name.
  5. Under Split for training, eval, or both, choose how Klu should label the selected items:
    • Both makes the items available to either workflow.
    • Training or Evaluation restricts the items to that split.
    • Optimize (80/20) assigns an 80% training and 20% evaluation split.
  6. To include every row that matches the current filters, enable Select all data (with current filter).
  7. Select Save Dataset.

Klu creates the Dataset and opens its detail page. The saved filter controls which Logs are included at creation time; the Dataset is then managed as its own collection.

Create or extend a Dataset by import

Open your App, select Datasets, and then select Import Data. If you start from the Datasets list, the import creates a new Dataset. If you select Import Data from an existing Dataset, it appends items to that Dataset.

Import a local file

  1. Select the Local File tab.
  2. When creating a Dataset, enter a Dataset Name. Klu uses the selected filename when the name is blank.
  3. Choose one .csv or .jsonl file.
  4. Set the Split to Both, Training, Evaluation, or Optimize (80% training/20% evaluation).
  5. Select Save.

The upload and import run asynchronously. The Dataset shows import progress, then refreshes its items and counts. An empty, malformed, or unsupported file ends with Import failed; correct the file and import it again.

Import an OpenAI file

If your Workspace has OpenAI files available, select the OpenAI File tab, choose an OpenAI File, set the Split, and select Save. Klu imports the selected provider file into a new or existing Dataset.

Add an item manually

The Manual tab appears when you open Import Data from an existing Dataset.

  1. Enter an Input as text or JSON.
  2. Enter the expected Output.
  3. Choose a Split.
  4. Select Save.

The new row becomes a Dataset item. Open any row to inspect its input, output, context, split, tags, token counts, and source. Use the Edit view in the item drawer to change its content, split, or tags.

Inputs can also represent prompt variables, message lists, and tools. Keep their structure compatible with the Action you plan to evaluate; otherwise an Eval may be unable to bind the Dataset input to that Action's prompt.

Manage Dataset items

The Dataset grid supports quick filtering and multi-row selection. The bulk-action controls let you:

  • Update tags or the training/evaluation split.
  • Relabel inputs or outputs with an LLM.
  • Generate additional examples from selected items.
  • Remove near-duplicates.
  • Export selected data.
  • Delete selected items.

Review generated or relabeled content before using it as evaluation truth or training data. These operations can change the meaning and distribution of the Dataset.

The Datasets list shows each Dataset's item count, token count, creator, last update, and associated fine-tuned model when one exists. You can rename a Dataset from its detail-page heading.

Use a Dataset in Evaluate

From a Dataset detail page, select Add Eval, choose the Action to evaluate, and complete the Eval setup. Klu uses the Dataset's evaluation-eligible items as test cases.

Use evaluation splits for held-out examples. If every item is marked Training, the Dataset does not provide a meaningful held-out set. See Evaluate prompt and model performance for evaluator setup, runs, and comparisons.

Use a Dataset in Fine-tune

Select Fine-tune from the Dataset detail page to start a training job with that Dataset. Klu requires more than 10 Dataset items before this action is available. The form also requires a connected provider that supports fine-tuning, a base model, and a fine-tune name.

Use training-eligible items for the job, and keep evaluation examples separate when you need an unbiased before-and-after comparison. See Fine-tune models for provider and model configuration.

Dataset API

Dataset API operations are separate from the UI workflow. The public SDKs can create and list Datasets, retrieve a Dataset by GUID, retrieve its data, add existing data GUIDs, and delete a Dataset. See SDK exports for the current clients and API basics for authentication and request conventions.