bert-squadv2-biomed

This model is a fine-tuned version of microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext on the SQuADv2 dataset. It has been fine-tuned for question-answering tasks specifically related to biomedical texts, leveraging the SQuAD v2 dataset to enhance its ability to manage both answerable and unanswerable questions.

Model Description

The base model, PubMedBERT, was originally pre-trained on biomedical abstracts and full-text articles from PubMed. This fine-tuned version adapts PubMedBERT for biomedical question-answering by training it with SQuADv2, a dataset that includes over 100,000 questions with answerable and unanswerable queries.

• Use Cases: This model is particularly useful in applications where quick and accurate question-answering from biomedical literature is needed. It is designed to provide answers to specific questions, as well as to detect when no relevant answer exists.

Training and Evaluation Data

• Dataset: The model was fine-tuned on the SQuADv2 dataset, which consists of reading comprehension tasks where some questions have no answer in the provided context.
• Training Environment: The model was trained in a Colab environment. A link to the training notebook can be found here: Training Notebook.

Training Procedure

Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam (betas=(0.9, 0.999), epsilon=1e-08)
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Framework versions

• Transformers 4.34.1
• Pytorch 2.1.0+cu118
• Datasets 2.14.5
• Tokenizers 0.14.1