risenyard/egms-qa-dataset

EGMS-QA Dataset

Natural-language question–answer records over European Ground Motion Service (EGMS) displacement time series. The release includes the prepared tiles, encoder tokens, labels, reference tables, and metadata behind those records.

Use the data

Read the published records with the QA example below. For code installation, label reconstruction, and new QA generation, use the GitHub QA construction guide. The Encoder and Translator repositories provide the associated model weights and recipes.

QA use

Install a current version of datasets:

pip install -U datasets

Read a QA record without downloading the full release:

from datasets import load_dataset

qa = load_dataset("risenyard/egms-qa-dataset", streaming=True)
print(next(iter(qa["test"])))

The default configuration exposes the three QA splits. Streaming reads records on demand. Omit streaming=True to download the QA splits. Source tiles and model artifacts are separate from this QA loading configuration.

QA records

HF split	record split value	tiles	QA records
train	train	8,000	554,000
validation	val	1,000	68,200
test	test	1,000	68,200

Each record pairs a tile and task with a question phrasing and reference answer. Task definitions are listed in the GitHub task catalog.

field	meaning
id	record identifier combining task, tile, and phrasing
tile_id, split	source tile and its fixed split assignment
task, target_column	task identifier and corresponding reference-table column
question, answer	visible natural-language question and reference answer
answer_type	numeric, categorical, refusal, or missing
answer_value	underlying target value or marker; its type depends on answer_type
rendered_target_value, decision_value	formatted numeric value or categorical decision used for rendering
phrase_id	identifier of the selected question phrasing
training_target, target_format, response_instruction	supervision text and response-format metadata

The files in data/qa/ are the fixed published splits. The GitHub generator can render additional corpora from the released labels and approved phrasings.

Release layers

path	role
data/qa/	train, validation, and test QA JSONL files
artifacts/source_tiles/	10,000 model-ready 7 km tiles in NPZ format
artifacts/representations/	egms_tokens_10k.pt cache and metadata, with 65×256 tokens per tile
artifacts/labels/	target table and metadata for the token-dependent tasks
artifacts/reference_tables/	per-group reference tables and frozen S3 temporal inputs
metadata/	data configuration, normalization, splits, audits, and integrity records

The GitHub installation workflow downloads the full release and maps it to the runtime paths used by the code.

Source-tile contract

The NPZ files contain processed measurements from the EGMS Level-3 Ortho Vertical product for 2019–2023.

array	contents
coords [N,2]	EPSG:3035 easting and northing
time_series [N,294]	model-ready vertical displacement in millimeters
pid [N]	persistent-scatterer identifiers
bbox [4]	tile bounding box

Auxiliary fields include height, RMSE, mean velocity, acceleration, seasonality, and their uncertainty fields. The encoder consumes stored indices [0,294), which correspond exactly to [8,302) on the 304-step source-preparation axis. The data config preserves the source offset and six-day cadence.

The split is fixed at tile level. Tiles overlap, so a persistent-scatterer ID may occur in more than one tile.

Reference tables

Each task group has a <group>_final_table.csv under artifacts/reference_tables/<group>/. The task implementation index links the definitions and computational methods.

S3 also uses frozen temporal inputs and their metadata. These contain posterior trend order and changepoint probability estimated with BEAST and define the temporal inputs to the published S3 targets. The S3 method documents their provenance and the variability of newly fitted estimates.

Normalization contract

metadata/normalization.json records the displacement mean, standard deviation, and linear-detrended residual standard deviation. They were fitted on the 8,000 training tiles over the stored [0,294) window. The identical file is provided with the released encoder.

Integrity

metadata/files.sha256 lists checksums for the released files. It excludes itself, metadata/release_manifest.json, and the Hub-managed .gitattributes. The release manifest summarizes the inventory and record counts, while the tile manifest records tile paths and hashes. Verification commands are in the GitHub setup guide.

Provenance, terms, and limitations

Generated using European Union's Copernicus Land Monitoring Service information; https://doi.org/10.2909/4a14a29b-7db7-40e4-81ad-df3aa8dfbc6f

The EGMS measurements were selected, arranged into overlapping 7 km research tiles, and combined with EGMS-QA-created labels and questions. This is not an official EU, EEA, Copernicus, or EGMS product; no endorsement is implied.

The QA records, derived labels/tables, and token cache are CC-BY-4.0. The Copernicus-derived measurements retain the CLMS attribution and modification conditions. See DATA_TERMS.md and SOURCE_PROVENANCE.md.

The data describe measured deformation histories. They do not establish cause, forecast future motion, certify structural safety, or replace expert InSAR and geotechnical assessment.