NoeFlandre/osm-polygon-web-search

OSM Polygon Web Search

This public Hugging Face dataset contains one explicitly approved POC table at train.parquet, with one row per non-empty SAT sentence classified as relevant-only by the local LiquidAI/LFM2.5-2.6B model. It is designed for the Hugging Face Dataset Viewer: the table exposes polygon geometry, the exact Brave query, page URL, title, HTTP status, full page text parsed by Trafilatura, the SAT sentence, and the strict relevance result. The complete yes/no classification table remains on the Seagate project volume.

The rows cover nine query variants: V1 land cover, V2 land use, V3 vegetation, V4 terrain, V5 soil surface, V6 ecosystem, V7 physical geography, V8 buildings infrastructure, and V9 landscape environment.

The table includes these main fields:

Field	Purpose
polygon_geojson, centroid, and bounding-box fields	Polygon geometry and map context
polygon_name, country, osm_type, osm_id	OSM identity
query_keyword, query	Approved keyword phrase and exact Brave query used for the search
page_rank, title, page_url, http_status	Retrieved-page metadata
text, text_char_count	Full page text extracted by Trafilatura and its length
sentence, sentence_index, sentence_count	SAT sentence and its position within the page
sentence_model	segment-any-text/sat-3l-sm model identifier
relevance_label	Local LFM classification; every published row is yes
relevance_model	LiquidAI/LFM2.5-2.6B model identifier

The extracted evidence and criteria fields are intentionally omitted. The full parsed text remains available for inspection, while sentence is the primary downstream text unit. Before SAT segmentation, obvious Trafilatura scaffolding is removed from the segmentation input: headings, menu and call-to- action fragments, metadata, identifiers, symbols, and very short fragments. The raw parsed text column is retained unchanged for provenance and review.

The source PBF and local working artifacts remain local. The uploaded table contains only the PBF basename in source_pbf, not a local filesystem path, and does not contain raw HTML or the provider response. Full parsed page text is included explicitly for this POC so the results can be inspected in the Viewer; the sentence-level sentence field is the primary downstream text unit. The Apache-2.0 license applies to the project’s original code and documentation; OSM-derived fields and web-derived text retain their source-specific terms and should not be treated as Apache-2.0 content.