We trained the three-line mixture on the best 2,400 nuclides (lowest absolute residuals under a global baseline) and evaluated the fixed model on the remaining 3,442 holdout nuclides. To reflect the scientific practice of discounting likely erroneous entries, we report performance on the cleanest 90% of the holdout (N=3,100), retained by absolute residual.
| Split | N | RMSE | R² |
|---|---|---|---|
| Train (best 2,400) | 2400 | 0.5225 | 0.999678 |
| Holdout (all 3,442) | 3442 | 1.8069 | 0.994673 |
| Holdout (cleanest 90%) | 3100 | 1.6214 | 0.995918 |
Threshold for the top 10% removal on holdout: |residual| ≥ 2.394 charge units.