README.md

---
pretty_name: scrolls datasets
tags:
  - herculaneum
  - vesuvius-challenge
---

# scrolls datasets

Working datasets for the scrolls pipeline: label sets, pseudo-label rounds and curated bundles used to train surface and ink models.

## Layout

- `rounds/<round>/` one directory per labeling round: images, labels, split manifest, stats
- `bundles/<name>/` task-specific subsets for training and evaluation

## Round manifest

Every round carries a `manifest.json` with seeds, source artifact hashes and the split registry, so any checkpoint can name the exact data it saw.

## Publication terms

Kept private during the competition. If a prize is accepted, everything here is published under CC-BY-NC 4.0 as the prize requires; license metadata is added at release.