|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- generated_from_trainer |
|
|
metrics: |
|
|
- accuracy |
|
|
model-index: |
|
|
- name: bert-paper-classifier |
|
|
results: [] |
|
|
--- |
|
|
|
|
|
# bert-paper-classifier |
|
|
|
|
|
This model is a fine-tuned version of [microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract](https://huggingface.co/microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract) on the dataset from [González-Márquez et al., 2023](https://www.biorxiv.org/content/10.1101/2023.04.10.536208v1). |
|
|
|
|
|
## Intended uses & limitations |
|
|
|
|
|
This model is intended to predict the category given the paper title (and optionally its abstract) — for the biomedical papers. For example, it is likely to predict `virology` as a category for the paper with a title containing `COVID-19`. |
|
|
|
|
|
So far only a subset of the PubMed dataset has been used for training. Future improvements to this model can come with using the full dataset with a combination of titles and abstracts for the fine-tuning as well as extending the training set to the preprints from bioRxiv and/or arXiv. |
|
|
|
|
|
## Training procedure |
|
|
|
|
|
The code for the model fine-tuning can be found [in the respective notebook](https://huggingface.co/oracat/bert-paper-classifier/blob/main/finetuning-pubmed.ipynb). |
|
|
|
|
|
### Training hyperparameters |
|
|
|
|
|
The following hyperparameters were used during training: |
|
|
- learning_rate: 5e-05 |
|
|
- train_batch_size: 128 |
|
|
- 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 |
|
|
|
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- Transformers 4.28.1 |
|
|
- Pytorch 2.0.0+cu117 |
|
|
- Datasets 2.11.0 |
|
|
- Tokenizers 0.13.3 |
