weknowinc/latam-real-estate-listings
HF LatAm Real Estate Dataset (v5) Export of listed real-estate properties from Costa Rica and Panamá for Hugging Face consumption (NLP, modeling, search, recommendation, price prediction, etc.). Version: v5 Sources: OpenSearch properties — Costa Rica (CR) OpenSearch pa_properties — Panamá (PA) Filters: status=listed, listing_date <= cutoff (default ~5 months old), must have price + (built or fallback) area Rows: one row per (property, operation) side. sale_and_rental expands to… See the full description on the dataset page: https://huggingface.co/datasets/weknowinc/latam-real-estate-listings.
HF LatAm Real Estate Dataset (v5)
Export of listed real-estate properties from Costa Rica and Panamá for Hugging Face consumption (NLP, modeling, search, recommendation, price prediction, etc.).
- Version:
v5 - Sources:
- OpenSearch
properties— Costa Rica (CR) - OpenSearch
pa_properties— Panamá (PA) - Filters: status=
listed,listing_date<= cutoff (default ~5 months old), must have price + (built or fallback) area - Rows: one row per (property, operation) side.
sale_and_rentalexpands to two rows with:venta/:alquilersuffixes inrecord_id. - record_id format:
{CR|PA}:{firestore_id}:{venta|alquiler} - Language: Spanish-first — descriptions, labels, operation types, and property types are in Spanish.
place_kinduses language-neutral snakecase machine codes. No English label columns (`placekindlabelen`, etc.) to avoid Spanglish. - Config source:
place_kindids, Spanish labels (label_es), and per-country defaults come fromapp/utils/geo_config.py(same as agentico-admin).
Place classification (place_kind)
Each row includes a standardized place type describing the kind of development or building context (e.g. vertical condominium, horizontal PH, office center, commercial plaza). This replaces legacy generic labels and is the primary field for grouping or featurizing listings by development type.
Exported fields:
Every row has a non-null place_kind valid for its country. Resolution order:
- Valid
place_kindon the property document - Linked complex (
condo_id) lookup in the condos index - Country default from
geo_config(default_place_kind)
Costa Rica (CR)
- Default (when listing has no explicit/linked classification):
condominio_horizontal— Condominio horizontal
Panamá (PA)
- Default (when listing has no explicit/linked classification):
ph_horizontal— PH horizontal
All columns
record_idcountryadmin_level_1,admin_level_2,admin_level_3place_kind,place_kind_label_es,place_kind_categoriahuman_verifiedproperty_type,property_subtypeoperation_type(venta|alquiler)price,currency,price_usd,price_per_m2_usdbuilt_area_m2,lot_area_m2bedrooms,bathrooms,parking_spaces,construction_yearconditionamenities,features(arrays, may be empty)description_mdlisting_month
Other column notes
admin_level_*: geographic admin names (province / canton or district / district or corregimiento).property_type/property_subtype: canonical Spanish housing types (e.g.apartamento,casa).human_verified:Truewhen the listing was touched in backoffice (admin_updatedon source).
Validation at export time
- Exact column set and order (
HF_COLUMNS) - No duplicate
record_id - Valid
operation_typeandcountry - Every row: non-null
place_kind+place_kind_label_es, allowed ids and Spanish labels checked againstapp/utils/geo_config.py
Generation
python scripts/export_hf_latam_dataset.py --version v5 --dry-run --limit 200
python scripts/export_hf_latam_dataset.py --version v5 --output-dir output/hf_latam_v5
python scripts/export_hf_latam_dataset.py --version v5 --output-dir output/hf_latam_v5 --jsonl
python scripts/check_hf_latam_export.py --parquet output/hf_latam_v5/train.parquetAlso writes export_stats.json alongside this README.
Stats from this run
- Version: v5
- Run date: 2026-06-21
- Cutoff month: 2026-01
- Output rows: 10394
- By country: {'Costa Rica': 8955, 'Panamá': 1439}
- By operation: {'venta': 8465, 'alquiler': 1929}
- By placekind: {'condominiohorizontal': 8942, 'condominiovertical': 13, 'phhorizontal': 1415, 'ph_vertical': 24}
See export_stats.json for skips, null rates, and sample record IDs.
License / usage
Intended for research and model training. Verify license of original listings before redistribution.
