> ## Documentation Index
> Fetch the complete documentation index at: https://docs.blindsight.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Data Playground

> Interactively explore, edit, slice, and re‑label samples before committing to a full scan.

The Data Playground is an interactive exploration workspace. It loads a
dataset (or subset) into an in‑memory **session** where you can poke
at the data, run lightweight previews, and publish the result back as
a new dataset or branch.

It's behind the `playground.access` permission, see
[Team & Access](/data-integrity/team).

## When to use it

* You just imported a dataset and want a feel for it before kicking
  off heavy scans.
* You need to **edit labels by hand** on a small subset to seed a
  cleaner training run.
* You want to **propose a stratified split** and visualize the class
  balance before saving it.
* You want a **quick bias peek** without paying for the full
  `bias_shortcut` engine.

## Sessions

A session is one user's working copy of a dataset, isolated from
everyone else. Sessions live on the Playground home page
(`/playground`).

| State       | What it means                                                   |
| ----------- | --------------------------------------------------------------- |
| `preparing` | Blindsight is materialising the working copy. Progress streams. |
| `ready`     | The Playground editor is open and editable.                     |
| `cancelled` | You cancelled preparation; the session is gone.                 |
| `failed`    | Preparation errored. Retry from the home page.                  |
| `idle`      | No edits for a while; eligible for auto‑cleanup.                |

Each session has its own isolated workspace so multiple teammates can
explore the same dataset in parallel without stepping on each other.

## Creating a session

<Steps>
  <Step title="Pick a source">
    From `/playground` click **New session**. Pick one of:

    * An existing dataset (whole, or filtered).
    * A new upload (zip / folder).
    * A pull from HuggingFace, Kaggle, or S3.
  </Step>

  <Step title="Wait for preparation">
    The session shows a streamed progress bar. You can cancel from
    the session card.
  </Step>

  <Step title="Open the editor">
    Once ready, the Playground opens with the data table, filters,
    and the analysis panels.
  </Step>
</Steps>

## What the editor gives you

| Pane               | What you can do                                                                          |
| ------------------ | ---------------------------------------------------------------------------------------- |
| **Data table**     | Paginate, sort, filter, inline‑edit cells. Multi‑select supports bulk label flips.       |
| **Filters**        | By class, split, file size, metadata field.                                              |
| **Embedding view** | Run an embedding pass and view a 2D projection. Lasso‑select points to filter the table. |
| **Bias preview**   | Lightweight version of the bias engine. Directional, not authoritative.                  |
| **Distributions**  | Class balance, file size, aspect ratio, token length, missing value counts.              |
| **Splits**         | Propose stratified / random / group splits and visualize the result before committing.   |

The bias preview and embedding view are **directional**, they're
meant to point you at problems, not replace a full scan. Treat them
like a sanity check.

## Saving your work

You have three ways to commit changes:

| Action                 | What it does                                               |
| ---------------------- | ---------------------------------------------------------- |
| **Publish as dataset** | Creates a new named dataset from the working copy.         |
| **Export CSV**         | Downloads the edited manifest. You can re‑import it later. |
| **Update cells**       | Commits edits back to a source dataset's branch.           |

Published datasets get a lineage edge `playground_session` linking
them to the session they came from, so auditors can trace any
hand‑edited samples back to who edited them.

## Common workflows

<AccordionGroup>
  <Accordion title="Quickly understand a new dataset">
    1. Create a session against the dataset, no subsetting.
    2. Open the embedding view; look for tight unexpected clusters.
    3. Skim the distributions panel for class imbalance or weird
       size outliers.
    4. Close the session; no need to publish.
  </Accordion>

  <Accordion title="Hand‑relabel a small slice">
    1. Filter the table to the class you want to clean.
    2. Use multi‑select + bulk label flip on visually mislabeled
       samples.
    3. Publish as a new branch on the original dataset.
    4. Train a quick model on the cured branch as an A/B test.
  </Accordion>

  <Accordion title="Propose a train / val / test split">
    1. Open the splits panel and choose **stratified by class**.
    2. Eyeball the proposed sizes per class.
    3. Adjust the random seed if a class is too small in a split.
    4. **Update cells** to commit the split column back to the
       dataset's branch.
  </Accordion>
</AccordionGroup>

## Sessions and cleanup

* Long‑idle sessions time out and are cleaned up automatically. The
  exact threshold is set on your tenant (typically 24h of inactivity).
* You can **rename**, **cancel**, **retry**, or **delete** any
  session from the Playground home page.
* Deleting a session does not affect the dataset it was sourced from.