---
pretty_name: ProbeScout Main17 tasks
task_categories:
  - image-classification
language:
  - en
tags:
  - probescout
  - visual-analytics
  - attribute-probes
---

# ProbeScout Main17 task packages

Use with [ProbeScout source and setup instructions](https://github.com/lyfdy456/ProbeScout).
Download only the dataset(s) you need, and extract each ZIP into the **code repository root**.
The ZIP paths already include `dataset/` and `visual_analytics/`.

| File | Tasks | Compressed size |
|---|---:|---:|
| cars.zip | 7 | 93.74 MB |
| hico.zip | 8 | 320.29 MB |
| celeba.zip | 2 | 439.43 MB |

Each package includes task definitions, attributes, query image IDs, ordered
records, original VQA source/fit/Validation labels, fixed partition contracts,
Web score/rank/projection/cluster arrays and the pinned initial F0 evidence.
There are **no original images, query pictures, thumbnails, historical human
feedback, user sessions or human-refined models**. Images are supplied locally
by the user from the original dataset distributions under their respective terms.

`task_packages.json` records ZIP and extracted manifest SHA-256 values.
Each extracted `dataset/tasks/<dataset>/manifest.json` records every payload hash.
The source release pins the immutable HF commit and manifest identities.
Original Val document hashes inside `validationProvenance` identify the original
training inputs, not the sanitized exported files; the latter have separate hashes.
This preserves compatibility with the published immutable probe banks.

After extracting and placing local images as documented:

```sh
uv run --project probe_learning python scripts/prepare_web.py --dataset cars
cd visual_analytics/pcp_analyze/web
npm run dev
```

Use `--dataset hico`, `--dataset celeba`, or multiple dataset names as appropriate.
`--without-images` prepares numerical views/training but cannot display photographs.
Cached browsing and weight refinement use the task package; training or native
probe updates additionally require [features](https://huggingface.co/datasets/Ian100/ProbeScout-features)
and, for updates, [pretrained banks](https://huggingface.co/Ian100/ProbeScout-probes).
Alternatively extract your own matching SigLIP features using the source scripts.
The original Cars, HICO-DET and CelebA dataset terms still apply.