> ## 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.

# Datasets

> Upload from anywhere, version with branches, trace lineage, and explore samples before kicking off a scan.

A dataset in Blindsight is a versioned collection of files (images,
text, or 3D volumes) plus its metadata and full lineage. Every scan,
every healing run, and every compliance report ties back to a specific
dataset version.

## Creating a dataset

From **Datasets → New**, pick a source. Each source captures
credentials once and remembers them for next time.

<Tabs>
  <Tab title="Local upload">
    | Source            | Accepts                                                           |
    | ----------------- | ----------------------------------------------------------------- |
    | **Upload ZIP**    | A zipped folder tree. Class name is taken from the parent folder. |
    | **Upload folder** | Drag‑drop a directory from your machine. Same class convention.   |
    | **Paste text**    | Plain‑text corpus, treated as a single text document.             |

    Blindsight does Zip Slip protection on extraction, so a malformed
    archive cannot escape the dataset directory.
  </Tab>

  <Tab title="Code / model hubs">
    | Source              | What you provide                                          |
    | ------------------- | --------------------------------------------------------- |
    | **GitHub repo**     | Repo URL, branch, optional path filter, token if private. |
    | **HuggingFace hub** | Repo id, optional config / split, optional sample cap.    |
    | **Kaggle dataset**  | Kaggle slug. Uses workspace‑level Kaggle API credentials. |

    HuggingFace imports honor 429 rate limits with exponential
    backoff. The retry counter is visible on the import card; rotate
    the token in **Settings → Integrations** if retries keep failing.
  </Tab>

  <Tab title="Object storage">
    | Source            | What you provide                                                   |
    | ----------------- | ------------------------------------------------------------------ |
    | **S3**            | Bucket + prefix. Uses the project's or workspace's S3 integration. |
    | **S3‑compatible** | Same shape, against MinIO, Wasabi, R2, Backblaze, etc.             |

    Use the **Test** button on the S3 integration card before kicking
    off a large import to confirm the credentials and the bucket
    region.
  </Tab>

  <Tab title="From a ticket">
    | Source               | What you provide                                             |
    | -------------------- | ------------------------------------------------------------ |
    | **Jira attachments** | Issue key. Pulls every attachment on the issue as a dataset. |

    Useful when an engineer files a "dataset looks wrong" ticket: you
    can import the attached samples and run a scan straight away.
  </Tab>
</Tabs>

## Dataset detail page (`/datasets/:id`)

The dataset page is your home base for one piece of data.

| Tab / section      | What it does                                                                                     |
| ------------------ | ------------------------------------------------------------------------------------------------ |
| **Preview grid**   | Lazy‑loaded thumbnails for image datasets, zoom on hover, class filter.                          |
| **File tree**      | The dataset's files with size, type, and class. Right‑click for actions (rename, move, exclude). |
| **Action history** | Every upload, scan, healing, branch, and edit, attributed to a user, timestamped, exportable.    |
| **Branches**       | Virtual versions of the dataset, each backed by a manifest CSV.                                  |
| **Lineage**        | Graph of parent / child datasets produced by healing, branching, or imports.                     |
| **Scan history**   | Past scans against this dataset, with severity and duration.                                     |
| **Scheduler**      | Assign a cron schedule. See [Automation](/data-integrity/automation).                            |
| **Settings**       | Rename, move between projects, configure webhooks, change webhook URL, delete.                   |

<Tip>
  Click any thumbnail in the preview grid to open a full‑size viewer
  with class metadata, EXIF, and the list of every scan that has
  touched that file. Useful while triaging a single weird sample.
</Tip>

## Branches

Branches let you maintain alternative versions of the same dataset
without physically copying files. Each branch is a manifest CSV that
declares which files are in or out, plus any label overrides.

**Typical uses:**

* A **clean** branch after healing, so you can A/B train on cured vs
  raw data without losing the original.
* A **balanced** branch with a downsampled subset for fast iteration.
* An **experimental** branch with tweaked labels for an A/B scan.

Switching branches on the dataset detail page instantly changes which
files are used in the next scan. Every branch keeps its own scan
history.

<Steps>
  <Step title="Create a branch">
    From the **Branches** tab, click **New branch**, give it a name,
    optionally describe how it differs from the main branch.
  </Step>

  <Step title="Define what's in it">
    Either upload a manifest CSV (`path,split,label,include`) or use
    the inline editor to flip the include flag on individual files.
  </Step>

  <Step title="Run a scan against it">
    The branch shows as a selector on the **New scan** screen. Pick
    it and Blindsight scans only the branch's manifest.
  </Step>

  <Step title="Promote or discard">
    Branches can be **promoted** to become the dataset's main view, or
    **deleted** without affecting any other branch.
  </Step>
</Steps>

## Lineage

The lineage view is a directed graph showing where a dataset came from
and what was produced from it.

* Healing produces a child dataset with a `healed_from` edge.
* Branching produces a child with a `branched_from` edge.
* Playground publishes produce a child with a `playground_session`
  edge.
* Imports keep a reference to the original `source_uri`.

Auditors typically ask "where did this training corpus come from?" and
lineage is the single answer. Compliance reports include the same
graph automatically.

## What to expect

* Dataset uploads under \~5 GB usually appear within seconds.
* Larger uploads stream progress via the WebSocket and resume
  cleanly if the network drops.
* Statistics (`valid_images`, `invalid_images`, `classes`,
  `class_count`) populate as soon as ingestion finishes; you don't
  have to wait for a scan to see them.