Fig. 3. The kernel density estimation curve of ORD yield
2.2.3 Labels smoothing
To address the limitation of traditional binary label encoding in
describing reaction superiority within the same label category, label
smoothing methods are employed to differentiate between varying degrees
of reaction superiority and reduce overfitting during model training. In
this case, the superiority labels for 20,000 reactions in the
fine-tuning dataset are reconstructed based on the assumption that the
overall effects of reaction yield, time and temperature conditions are
the uniform and linear. According to the compensation part equations
provided in Table 3 , the high superiority reactions labels are
remapped to a range of 0.85 to 1.0, and the low superiority reaction
labels are remapped to a range of 0.0 to 0.15.