Upload QUALITY_RUBRIC.md with huggingface_hub

QUALITY_RUBRIC.md

@@ -0,0 +1,78 @@
+# Quality Score Rubric
+
+## Overview
+
+Each entry receives a composite quality score (0–88) from five weighted sub-scores.
+No entry scores >= 90 by design — the scoring system is conservative and penalizes any missing information.
+
+## Sub-Score Breakdown
+
+| Sub-score | Max | Mean | Weight | Description |
+|-----------|-----|------|--------|-------------|
+| Geometry | 23 | 16.7 | 23% | Structure validity, volume presence, nsites sanity, lattice parameter plausibility |
+| DFT | 20 | 19.7 | 20% | Formation energy and band gap completeness, physical range checks |
+| Metadata | 15 | 11.8 | 15% | Space group, density, crystal system, provenance completeness |
+| Novelty | 15 | 14.2 | 15% | Cross-source deduplication status, duplicate group membership |
+| Chemical | 15 | 15.0 | 15% | Element validity, composition plausibility, charge balance |
+
+## Scoring Rules
+
+### Geometry (max 23)
+- +5: structure_json present and parseable
+- +5: volume > 0
+- +5: nsites > 1 (penalize single-atom entries)
+- +4: lattice parameters within physical ranges (a,b,c > 2Å, angles 0-180°)
+- +4: density > 0 and < 50 g/cm³
+
+### DFT (max 20)
+- +5: formation_energy_per_atom present
+- +5: formation_energy_per_atom in [-5, 5] eV/atom (penalize extreme outliers)
+- +5: band_gap present
+- +5: energy_above_hull present
+
+### Metadata (max 15)
+- +5: space_group present
+- +5: density present and > 0
+- +3: crystal_system present
+- +2: provenance fields complete
+
+### Novelty (max 15)
+- +15: unique entry (no duplicate across sources)
+- +10: member of duplicate group (penalty for redundancy)
+- Deduction depends on number of duplicates
+
+### Chemical (max 15)
+- +5: valid elements (all known periodic table elements)
+- +5: plausible composition (positive atom counts)
+- +5: no unphysical stoichiometries
+
+## Score Bins and Interpretation
+
+| Score Range | Count | Interpretation |
+|-------------|-------|----------------|
+| 80–88 | 83,649 | High quality: all critical fields present, no outliers |
+| 70–79 | 146,001 | Good quality: minor gaps in metadata or geometry |
+| 60–69 | 32,273 | Fair quality: notable gaps in structure/metadata |
+| 50–59 | 4,807 | Low quality: missing multiple critical fields |
+| 40–49 | 2 | Failed most checks (retained for debugging) |
+
+## Calibration Evidence
+
+Higher score bins show strictly better metrics:
+
+| Score Bin | Valid Structure % | Space Group % | Outlier FE % |
+|-----------|---|------|---|
+| 50–60 | 0.1% | 99.8% | 0.31% |
+| 60–70 | 27.9% | 2.0% | 0.01% |
+| 70–80 | 77.4% | 8.8% | 0.00% |
+| 80–90 | 85.2% | 97.4% | 0.00% |
+
+See `dataset/statistics/quality_calibration.json` for full calibration data.
+
+## Known Limitation: MP vs OQMD Score Bias
+
+- MP entries: mean score **87.4**
+- OQMD entries: mean score **73.2**
+
+This reflects genuine quality differences (MP uses finer computational parameters), not systematic bias.
+Use tier-based filtering (Gold/Validated/Raw) to isolate quality levels across sources.