raynardj/roberta-pubmed

Roberta-Base fine-tuned on PubMed Abstract

We limit the training textual data to the following MeSH

• All the child MeSH of Biomarkers, Tumor(D014408), including things like Carcinoembryonic Antigen(D002272)
• All the child MeSH of Carcinoma(D002277), including things like all kinds of carcinoma: like Carcinoma, Lewis Lung(D018827) etc. around 80 kinds of carcinoma
• All the child MeSH of Clinical Trial(D016439)
• The training text file amounts to 531Mb

Training

• Trained on language modeling task, with mlm_probability=0.15, on 2 Tesla V100 32G

training_args = TrainingArguments(
    output_dir=config.save, #select model path for checkpoint
    overwrite_output_dir=True,
    num_train_epochs=3,
    per_device_train_batch_size=30,
    per_device_eval_batch_size=60,
    evaluation_strategy= 'steps',
    save_total_limit=2,
    eval_steps=250,
    metric_for_best_model='eval_loss',
    greater_is_better=False,
    load_best_model_at_end =True,
    prediction_loss_only=True,
    report_to = "none")