This
ChemBERTa-v2 checkpoint was fine-tuned on the
USPTO-50k dataset for sequence classification.
Specifically, the objective is to predict the reaction class label, and the input is either (canonicalized) all reactant SMILES or all product SMILES (separated by ".").
-
Train/Test split: 0.99/0.01
-
Evaluation results:
- Accuracy: 87.11%
- Loss: 0.4272
-
Fine-tuning hyperparameters:
- seed = 233
- batch-size = 128
- num_epochs = 5 (but early stopped at epoch 4)
- learning_rate = 5e-4
- warmup_steps = 64
- weight_decay = 0.01
- lr_scheduler_type = "cosine"
