license: mit
base_model: Sarmila/pubmed-bert-squad-covidqa
tags:
- generated_from_trainer
- biology
datasets:
- covid_qa_deepset
- squad
model-index:
- name: pubmed-bert-squad-covidqa
results: []
language:
- en
pipeline_tag: question-answering
pubmed-bert-squad-covidqa
This model is a fine-tuned version of microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext on the squad qa first, covid_qa_deepset dataset. It achieves the following results on the evaluation set for squad: {'exact_match': 59.0, 'f1': 76.32473929579194}
- Loss 1.003116
It achieves the following results on the evaluation set for covidqa:
- Loss: 0.4876
Model description
This model is trained with an intention of testing pumed bert bionlp language model for question answering pipeline. While testing on our custom dataset, we reliazed that the model when used directly for QA did not perform well at all. Hence, we decided to train on covidqa to make model accustomed with answer extraction. While, covidqa data is very similar to what we intended to use, it is samll in number hence resulting not much improvement.
Therefore, we firt trained the model in squad dataset which is larger in number. Then, we trained the model for covid qa. Hence, squad helped model to learn how to extract answers and covid qa helped us to train the model on domain similar to ours i.e. biomedicine
further, we have first performed MLM using our dataset on pubmed bert bionlp and then performed exactly same 眉i眉eline to see the difference which is [here]
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: 2e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| No log | 1.0 | 51 | 0.4001 |
| No log | 2.0 | 102 | 0.4524 |
| No log | 3.0 | 153 | 0.4876 |
Framework versions
- Transformers 4.33.0
- Pytorch 2.0.0
- Datasets 2.1.0
- Tokenizers 0.13.3