MINDSET

MINDSET is the pretraining dataset for MIND, a coordinate-only location encoder distilled from static location encoder teachers and annual AlphaEarth Foundations (AEF) embeddings.

We release the embeddings at the 12.1M training coordinates. The dataset contains 12,099,072 land coordinates in WGS84. Coordinates are dense around cities and not uniformly sampled over land.

The files are in GeoParquet format and can be joined on point_id:

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

Coordinate and array conventions

• Geometry is OGC:CRS84: Point(lon, lat) in WGS84 degrees.
• The coordinate input used by MIND is ordered (lat, lon); the GeoParquet geometry is ordered (lon, lat).
• The static teacher arrays are raw float16 embedder outputs. L2-normalize them when constructing targets.
• AEF is stored as native signed int8. For valid values, dequantize with f = sign(x) * (|x| / 127.5) ** 2; -128 is nodata. L2-normalize after dequantization.
• AEF covers years 2017--2025.

Teachers

column	teacher	dimension	source
aef	AlphaEarth Foundations	64	Google AEF v1 annual, source.coop tge-labs/aef
climplicit	Climplicit	1024	Jobedo/climplicit
geoclip	GeoCLIP	512	GeoCLIP location encoder
sinr	SINR	256	MVRL/sinr-location-encoder-1000-cls

The dataset contains CC-BY-4.0 metadata. The files contain embeddings derived from upstream models, so the license and use terms of each upstream model and source apply.

Load

import pyarrow.dataset as ds
import huggingface_hub
import fsspec

teachers = ds.dataset(
    "hf://datasets/XXXX/MINDSET/mindset_teachers.parquet", format="parquet"
)
aef = ds.dataset(
    "hf://datasets/XXXX/MINDSET/mindset_aef.parquet", format="parquet"
)
# Scan only columns and rows needed for a spatial AOI
sample = teachers.to_table(columns=["point_id", "geometry"], filter=(ds.field("point_id") < 10))