README.md

metadata

license: cc-by-4.0
task_categories:
  - feature-extraction
tags:
  - geospatial
  - location-encoder
  - embeddings
  - distillation
  - geoparquet
pretty_name: MINDSET
size_categories:
  - 10M<n<100M

MINDSET — MIND Spatial-Embedding Teachers

The pretraining dataset for MIND (Matryoshka Implicit Neural Distillation), a lat/lon-only location encoder distilled from four geospatial teachers. MINDSET caches the teachers' location-conditioned embeddings so the pretraining is reproducible and the location-only teachers do not need to be recomputed each epoch.

Points are the urban-dense global training sample (12,099,072 land coordinates, WGS84). Two files, joinable on point_id, both GeoParquet 1.1 (Hilbert-sorted, bbox covering column → spatial range queries read only the relevant row groups):

file	grain	rows	columns
mindset_teachers.parquet	one row per point	12,099,072	point_id, geometry (Point), bbox, climplicit [1024], geoclip [512], sinr [256] — float16
mindset_aef.parquet	one row per (point, year)	108,891,648	point_id, year, geometry, bbox, aef [64] — int8

Conventions

• CRS: OGC:CRS84 (lon, lat) WGS84 — geometry is Point(lon, lat). (Not EPSG:4326, which is lat,lon.)
• Join the two files on point_id. The three teachers are location-only (one vector per point); AlphaEarth is temporal, so it has one row per (point, year), years 2017–2025.
• dtypes / dequantization. Teachers are stored raw as float16 (the embedder output). AEF is the native int8 with signed_square quantization — dequantize with f = sign(x) * (|x| / 127.5) ** 2 (nodata = −128). Both are turned into training targets by L2-normalizing at use (MIND also optionally PHI-S-standardizes the teachers). float16/int8 are downstream-lossless for distillation targets.

Teachers (provenance)

column	teacher	dim	source
aef	AlphaEarth Foundations	64	Google AEF v1 annual (source.coop tge-labs/aef), CC-BY-4.0
climplicit	Climplicit	1024	Jobedo/climplicit (CHELSA climate, ReSIREN)
geoclip	GeoCLIP	512	GeoCLIP location encoder
sinr	SINR	256	MVRL/sinr-location-encoder-1000-cls (Cole et al. 2023)

Embeddings are derived from these pretrained models; respect each upstream license.

Load

import geopandas as gpd, pandas as pd
teachers = gpd.read_parquet("mindset_teachers.parquet")  # geometry + climplicit/geoclip/sinr
aef = gpd.read_parquet("mindset_aef.parquet")            # per (point, year); join on point_id
# spatial subset without a full download (cloud-optimized): DuckDB spatial / pyarrow dataset filters

Reproduce

Built with scripts/build_mindset.py from the neuralftw repo (AEF from the existing int8 shards; the three location-only teachers computed on GPU).