Table 3 Comparison of model accuracy based on selected attention mechanism on combined datasets
From: Substructure-based neural machine translation for retrosynthetic prediction
|  | Maximum reactant–product pair length | |||
|---|---|---|---|---|
100 | 120 | 140 | 160 | |
Dataset size | 350 K | 489 K | 588 K | 637 K |
Testset size | 55 K | 79 K | 96 K | 105 K |
Global (\(T_{c} \ge 0.85\)) | 38.2% | 37.5% | 37.2% | 37.2% |
Local (\(T_{c} \ge 0.85\)) | 38.3% | 39.6% | 38.6% | 38.6% |