# sample/

**1 file**: `sample_map.jsonl` — a **17,911-polygon** spatial
sample of the full worldwide dataset, designed to be **highly
representative** of the geographic distribution without being too
dense to render in a browser. Use it as a quick look at "what does
this dataset look like?" without downloading the 13 GB combined
parquet. Schema is centroid-only (no geometry column).

Generated by `scripts/sample_for_map.py`. Algorithm:

1. **Power-law per-country allocation**: each country gets
   `clamp(round(n_polygons ** 0.4), 8, 200)` polygons. This compresses
   the per-country range so dense countries (germany: 1.1M) don't crowd
   out sparse ones (monaco: 2), while still preserving the relative
   density differences.
2. **Spatial coverage within each country**: each polygon's centroid is
   bucketed into a `K x K` grid (`K = ceil(sqrt(target_n))`), then one
   random polygon per cell is kept. This avoids clustering in dense
   regions and ensures the sampled polygons are well-spread.
3. **Floor of 8** so even monaco's 2 polygons are visible.
4. **Cap of 200** so germany doesn't fill the map.

The same sample powers [`../preview/map_preview.png`](../preview/),
which is a static PNG rendered via headless Chromium.

[Back to the dataset root](../README.md)