bert-base-uncased-finetuned-MAOB-IC50s-V2

This model is a fine-tuned version of bert-base-uncased on 4828 MAOB IC50 values from ChEMBL.

It achieves the following results on the evaluation set:

• Loss: 0.7730
• Accuracy: 0.7393
• F1: 0.7410

Use in Pipeline:

from transformers import pipeline

ic50_pipe = pipeline("text-classification", model="cafierom/bert-base-uncased-finetuned-MAOB-IC50s-V2")

bert_ic50 = ic50_pipe('C#CCN(C)[C@H](C)Cc1ccccc1')

#result: [{'label': '< 50 nM', 'score': 0.7967771291732788}]

Model description

More information needed

Intended uses & limitations

Can classify MAOB IC50 values as < 50 nM, < 500 nM, and > 500 nM. See Confusion matrix below:

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• 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	Accuracy	F1
0.9383	1.0	129	0.9193	0.5848	0.4316
0.8509	2.0	258	0.7728	0.6745	0.6173
0.7629	3.0	387	0.7696	0.6607	0.6141
0.7104	4.0	516	0.7762	0.6455	0.6134
0.6468	5.0	645	0.6990	0.7103	0.6886
0.61	6.0	774	0.6950	0.7145	0.7029
0.5699	7.0	903	0.7075	0.6952	0.7021
0.5439	8.0	1032	0.6824	0.7269	0.7217
0.5185	9.0	1161	0.6768	0.7352	0.7328
0.4814	10.0	1290	0.6878	0.7172	0.7158
0.4665	11.0	1419	0.7202	0.7324	0.7187
0.4344	12.0	1548	0.7586	0.7034	0.7077
0.4209	13.0	1677	0.7469	0.7090	0.7114
0.3912	14.0	1806	0.7304	0.7324	0.7253
0.3828	15.0	1935	0.7357	0.7269	0.7308
0.3727	16.0	2064	0.7694	0.72	0.7254
0.3517	17.0	2193	0.7709	0.7393	0.7376
0.3406	18.0	2322	0.7667	0.7366	0.7386
0.3282	19.0	2451	0.7683	0.7297	0.7289
0.3161	20.0	2580	0.7730	0.7393	0.7410

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.3.0
• Tokenizers 0.21.0