pubmedbert-fulltext-cord19

This model is a fine-tuned version of microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext on the pritamdeka/cord-19-fulltext dataset. It achieves the following results on the evaluation set:

• Loss: 1.2667
• Accuracy: 0.7175

Model description

The model has been trained using a maximum train sample size of 300K and evaluation size of 25K due to GPU limitations

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 10000
• num_epochs: 3.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.7985	0.27	5000	1.2710	0.7176
1.7542	0.53	10000	1.3359	0.7070
1.7462	0.8	15000	1.3489	0.7034
1.8371	1.07	20000	1.4361	0.6891
1.7102	1.33	25000	1.3502	0.7039
1.6596	1.6	30000	1.3341	0.7065
1.6265	1.87	35000	1.3228	0.7087
1.605	2.13	40000	1.3079	0.7099
1.5731	2.4	45000	1.2986	0.7121
1.5602	2.67	50000	1.2929	0.7136
1.5447	2.93	55000	1.2875	0.7143

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.11.0