bert-base-cased-ChemTok-ZN15-55KStat-V1

This model is a fine-tuned version of bert-base-cased on on the cafierom/ZN1540K dataset of drug or drug-like molecules.

Model description

This domain adaptation of bert-base-cased has been trained on ~56.5K molecular SMILES strings, with added tokens:

new_tokens = ["[C@H]","[C@@H]","(F)","(Cl)","c1","c2","(O)","N#C","(=O)",
              "([N+]([O-])=O)","[O-]","(OC)","(C)","[NH3+]","(I)","[Na+]","C#N"]

Intended uses & limitations

It is meant to be used for finetuning classification models for drug-related tasks, and for generative unmasking.

Training and evaluation data

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 128
• eval_batch_size: 128
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss
1.7456	1.0	376	0.8446
0.701	2.0	752	0.5086
0.5099	3.0	1128	0.4165
0.4384	4.0	1504	0.3661
0.3883	5.0	1880	0.3302
0.361	6.0	2256	0.3102
0.3417	7.0	2632	0.2967
0.3211	8.0	3008	0.2734
0.3099	9.0	3384	0.2670
0.2998	10.0	3760	0.2592
0.2919	11.0	4136	0.2540
0.2829	12.0	4512	0.2430
0.2729	13.0	4888	0.2322
0.2661	14.0	5264	0.2336
0.2631	15.0	5640	0.2260
0.2578	16.0	6016	0.2293
0.2547	17.0	6392	0.2251
0.2515	18.0	6768	0.2206
0.2503	19.0	7144	0.2174
0.2512	20.0	7520	0.2169

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0