LANYUE-Bench

The seven newly constructed datasets accompanying LANYUE: A Lunar Foundation Model for Unified Representation Learning. The full benchmark has 12 datasets; the five datasets derived from public benchmark sources are obtained separately from their original authors.

Datasets

One sample means one multimodal tuple, registration pair or crater identity, rather than one file per observation source. This release contains 29,965 samples. Image dimensions below describe the stored files.

Dataset	Samples	Train / val / test	Observations and stored sizes	Target
MSLCD	8,881	6485 / 1401 / 995	CE-2 DOM/DEM 512×512; WAC 48×48; NASA DEM 64×64	Crater detection boxes
ODPLCD-Seg	2,494	1592 / 395 / 507	WAC / DEM, 593×593	Catalogue-derived crater footprints
LunarOSR	1,750 pairs	953 / — / 797	CE-2 DOM / Mini-RF, 512×512	7–11 control-point pairs per image pair
LunarICR	1,200 identities	840 / 120 / 240	CE-2 DOM / DEM / WAC / SLDEM, all 512×512	Cross-source crater identity
LunarSAC	14,634	10438 / 1490 / 2706	WAC / DEM, 512×512	Five surface-age groups
LunarSCC	600	420 / 90 / 90	Mini-RF CPR, 224×224	Older / younger catalogue model-age group
LunarPSRC	406	294 / — / 112	CE-2 DOM / WAC, 512×512	Non-PSR / PSR

An absent validation split is represented by an empty val.jsonl. Stored size is separate from model input size: LunarPSRC commonly uses 224×224 inputs. The sldem key denotes the NASA DEM branch called lola_dem or wac_dem in upstream code; the key alone does not identify a product version.

Download and load

The repository distributes ordinary TAR archives, JSONL manifests and a NumPy reader. It does not require converting GeoTIFFs or scientific arrays to RGB. This archive release does not provide automatic datasets.load_dataset() or the Hugging Face Dataset Viewer.

python -m pip install -U huggingface_hub
hf download LLukaoo/LANYUE-bench --repo-type dataset --local-dir /data/LANYUE-bench
cd /data/LANYUE-bench
sha256sum -c SHA256SUMS
mkdir -p data
for archive in archives/*.tar; do
  tar -xf "$archive" -C data
done
python -m pip install -r requirements.txt

Download a single dataset by specifying archives/LunarSAC.tar after the repo ID; also download benchmark/* and requirements.txt to use the supplied reader. For reproducible use, add --revision <commit-or-tag> to the download command.

from benchmark import BenchmarkDataset

dataset = BenchmarkDataset("/data/LANYUE-bench/data/LunarSAC", split="train")
sample = dataset[0]
print(sample["id"], sample["modalities"]["wac_dom"].shape, sample["target"])

The reader returns arrays in C,H,W order with stored dtype and values. Resizing, normalization and task-specific batching are the caller's responsibility.

File and label format

Each TAR extracts to <dataset-name>/ containing dataset.json, train.jsonl, val.jsonl, test.jsonl, checksums.sha256 and assets/. Manifest asset paths are relative to that dataset directory. Verify the original asset bytes with sha256sum -c checksums.sha256 from inside the extracted dataset directory.

• MSLCD: single-class YOLO rows [class, cx, cy, w, h]; class is 0 and boxes are normalized to the CE-2 DOM image. Four raster files belong to each sample.
• ODPLCD-Seg: mask values 0 background, 1 crater footprint, 255 ignore. Footprints are filled Robbins fitted ellipses, rather than manually traced pixel boundaries.
• LunarOSR: stored SAR GeoTIFFs have four bands; the reader selects band 0. Stored point rows are [x0, y0, x1, y1] with negative y values. The reader multiplies both y columns by −1 to return positive-down pixel coordinates.
• LunarICR: 1,200 distinct crater identities with four observations each (4,800 raster files). Query observations are CE-2 DOM/DEM; gallery observations are WAC/SLDEM. All observations of an identity remain in the same split.
• LunarSAC: 0 pre-Nectarian, 1 Nectarian, 2 Imbrian, 3 Eratosthenian, 4 Copernican. Labels are geological age groups, rather than absolute ages.
• LunarSCC: 0 older (model age ≥2.0 Ga), 1 younger (≤0.5 Ga); intermediate ages are excluded. Inputs are Float32 CPR (SC/OC) NPY arrays with validity masks. Invalid pixels were filled with the training valid-pixel mean 0.5427012881105048; the corresponding standard deviation is 0.22612217643426372. The reader does not apply z-score normalization.
• LunarPSRC: 0 non-PSR, 1 PSR; 290/116 samples respectively. Labels identify the selected PSR regions, rather than directly measuring water-ice presence.

Sources and license

Observations come from Chang'e-2, LROC WAC, NASA lunar elevation products and Mini-RF. Labels and identities use existing curated annotations and catalogues: Robbins crater geometry (ODPLCD-Seg and LunarICR), the USGS Unified Geologic Map (LunarSAC), corrected FT2014 model ages (LunarSCC), and Lemelin et al. PSR regions (LunarPSRC). See NOTICE and distributed per-sample metadata.

The supplied reader is MIT-licensed; see LICENSE. This code license does not automatically relicense third-party imagery, maps or catalogues. A dataset-wide license must be finalized by the publisher before publication; this draft card does not assert one.

Citation

Paper URL and BibTeX will be added with the paper release.