---
language:
  - en
tags:
  - probescout
  - visual-analytics
  - image-retrieval
---

# ProbeScout features

Frozen SigLIP features for Stanford Cars, HICO-DET and CelebA:
float32 global embeddings `[N, 768]`, float16 patch tokens `[N, 196, 768]`,
and records that map each row to an image ID. Total size: 81.10 GB.

Use the [ProbeScout setup guide](https://github.com/lyfdy456/ProbeScout/blob/codex/initial-source/docs/manual_setup.md)
and [dataset-specific download commands](https://github.com/lyfdy456/ProbeScout/blob/codex/initial-source/docs/asset_distribution.md).
Download only the dataset you need. Place its `dataset/` contents under the code
repository's `dataset/` directory, preserving paths and row order.

These features skip extraction for training and probe updates. Original images
are still required for photograph previews. Browsing saved tasks and using the
paper's Weight Tune / Staged feedback only require images and the
[task package](https://huggingface.co/datasets/Ian100/ProbeScout-tasks).

## CelebA patch tokens

The 60.99 GB CelebA patch NPY is stored as 29 byte-range parts. From the download
directory containing `restore_patch_features.py`, run:

```sh
python restore_patch_features.py --root .
```

Restoration verifies file hashes and needs an additional 60.99 GB of disk space.
It keeps the parts. Cars and HICO patch tokens are ordinary NPY files.

The code pins payload revision `3ef237a82d0c738a159aa49bb86764732d331265`.
Checksums are in `asset_manifest.json`. Original dataset terms apply;
photographs and historical human feedback are excluded.