SELFIES-bert-PubChem10M

This model is a scratch trained version of BERT on the PubChem10M dataset. It achieves the following results on the evaluation set:

Loss: 0.1290
Accuracy: 0.9544

Model description

BERT molecular language model. Trained on the Self-Referencing Embedded Strings (SELFIES) molecular representation. Tokenizer trained with a SELFIES semantically robust alphabet.

Intended uses & limitations

Used to extract embeddings of molecular representations, for downstream tasks (eg. binding affinity, toxicity prediction).

Training and evaluation data

Trained on a full split of the PubChem10M dataset, converted to SELFIES. Trained using run_mlm.py from the 🤗 examples, using torchrun for multi-GPU.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 32
eval_batch_size: 32
seed: 42
distributed_type: multi-GPU
num_devices: 2
total_train_batch_size: 64
total_eval_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3.0

Training results

The computations described in this research were performed using the Baskerville Tier 2 HPC service (https://www.baskerville.ac.uk/). Baskerville was funded by the EPSRC and UKRI through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure programme (EP/W032244/1) and is operated by Advanced Research Computing at the University of Birmingham.

ejy
/

SELFIES-BERT-PubChem10M