Datasets:

Scandium-Labs
/

Scandium-Dataset

Quality scores are not just "how good is this entry" — they measure trustworthiness for machine learning. An entry might be perfectly valid DFT data but score low on novelty (too many similar entries). This is intentional: a diverse, well-sampled dataset produces better models than one with 100,000 copies of Li₂O.

Score Components

The quality score is a composite of five sub-scores (max 95, normalized to 0-100):

Sub-score	Max	What it measures
Geometry	25	Structure quality: nsites, volume, density reasonableness
DFT	20	Property completeness and physical range
Metadata	15	Space group, density, structure_json presence
Novelty	20	How unique this entry is relative to the dataset (diversity bonus)
Chemical	15	Chemical reasonableness (element combinations, valence, etc.)

Calibration

We proved the quality score is monotonically related to label reliability:

Score Bin	N	Valid%	SG%	Complete%	FE Out%	FE MAE
50-60	4,807	0.1%	0.1%	0.0%	0.31%	—
60-70	32,273	27.9%	2.0%	0.0%	0.01%	0.80
70-80	146,001	77.4%	8.8%	0.1%	0.00%	0.23
80-90	83,649	85.2%	97.4%	82.7%	0.00%	0.18

Key finding: score 80-90 is research-grade:

97.4% have verified space groups
0% property outliers
0.18 eV/atom cross-source FE MAE (close to DFT noise floor)

Thresholds

Threshold	Purpose
≥ 70	Validated tier — basic research use
≥ 80	Strict Gold — publication-grade, benchmark-ready
≥ 90	Not yet achieved — reserved for experimentally validated entries

Score Distribution

50-60:   4,807 (1.8%)   ██
60-70:  32,273 (12.1%)  ████████████
70-80: 146,001 (54.7%)  ████████████████████████████████████████████████
80-90:  83,649 (31.4%)  ████████████████████████
90-100:  0 (0.0%)

Sub-score Means

Sub-score	Mean	Max
Geometry	16.7	25
DFT	19.7	20
Metadata	11.8	15
Novelty	14.2	20
Chemical	15.0	15

Limitations

No score ≥ 90 — scoring is conservative; no entry has "perfect" metadata
OQMD entries score lower on metadata (missing space group until computed)
Novelty scoring may unfairly penalize common structures