README.md

---
pretty_name: Map Annotation Tool Media
task_categories:
- image-classification
license: other
---

# Map Annotation Tool Media

This dataset contains Docling-extracted figure crops grouped by source PDF. It is the
media input for `map-annotation-tool`; human annotations are maintained separately in
the application repository and submitted through pull requests.

## Contents

- `media/figures/brgm-v1/`: 1,512 figures from the earlier BRGM parsing batch,
  of which 156 had legacy annotations at migration time.
- `media/figures/brgm-v2/`: 6,046 figures from the newer BRGM parsing batch.
- `figures.jsonl`: portable figure manifest with dimensions, content SHA-256, source
  document, relative media path, and MapPool prediction metadata.
- `annotations.jsonl`: migrated human ground-truth annotations when published.
- `mappool_predictions.csv`: portable map/not-map labels and decision margins.
- `build-report.json`: input/output counts, model provenance, and manifest checksum.

MapPool predictions are queue-prioritization hints, not ground truth. The classifier
uses the `sraimund/MapPool` head with the OpenCLIP
`ViT-L-14-quickgelu/openai` backbone. Neither model weight is included here.

Published human annotations include classification, map type, georeferencing status,
optional GCPs and geometry, contributor provenance, and timestamps. Annotation records
default to CC BY 4.0; that does not relicense the source figure media.

## Licensing and release status

This repository is a public preview so the media, provenance, attribution, checksums,
predictions, and annotation workflow can be reviewed end to end.

The original annotation tooling is MIT-licensed and annotation records default to
CC BY 4.0. Those licenses do **not** relicense the source figure media. Before changing
reusing this dataset, consult the applicable BRGM public-document statement and
attribution. MapPool predictions are distributed as metadata only; no MapPool or
OpenCLIP model weights are included.