PubChem-10m-t5

This model is a fine-tuned version of google/t5-v1_1-base on the sagawa/pubchem-10m-canonicalized dataset. It achieves the following results on the evaluation set:

• Loss: 0.2121
• Accuracy: 0.9259

Model description

We trained t5 on SMILES from PubChem using the task of masked-language modeling (MLM). Its tokenizer is also trained on PubChem.

Intended uses & limitations

This model can be used for the prediction of molecules' properties, reactions, or interactions with proteins by changing the way of finetuning.

Training and evaluation data

We downloaded PubChem data and canonicalized them using RDKit. Then, we dropped duplicates. The total number of data is 9999960, and they were randomly split into train:validation=10:1.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-03
• train_batch_size: 30
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 30.0

Training results

Training Loss	Step	Accuracy	Validation Loss
0.3866	25000	0.8830	0.3631
0.3352	50000	0.8996	0.3049
0.2834	75000	0.9057	0.2825
0.2685	100000	0.9099	0.2675
0.2591	125000	0.9124	0.2587
0.2620	150000	0.9144	0.2512
0.2806	175000	0.9161	0.2454
0.2468	200000	0.9179	0.2396
0.2669	225000	0.9194	0.2343
0.2611	250000	0.9210	0.2283
0.2346	275000	0.9226	0.2230
0.1972	300000	0.9238	0.2191
0.2344	325000	0.9250	0.2152
0.2164	350000	0.9259	0.2121