---
language: en
tags:
- science
- multi-disciplinary
license: apache-2.0
---

# ScholarBERT_100_64bit Model

This is the **ScholarBERT_100_64bit** variant of the ScholarBERT model family. The difference between this variant and the **ScholarBERT_100** model is that its tokenizer 
is trained with `int64` rather than the default `int32`, so the count of very frequent tokens (e.g., "the") does not overflow.

The model is pretrained on a large collection of scientific research articles (**221B tokens**).

This is a **cased** (case-sensitive) model. The tokenizer will not convert all inputs to lower-case by default.

The model is based on the same architecture as [BERT-large](https://huggingface.co/bert-large-cased) and has a total of 340M parameters.


# Model Architecture

| Hyperparameter  | Value |
|-----------------|:-------:|
| Layers     | 24    |
| Hidden Size     | 1024  |
| Attention Heads | 16    |
| Total Parameters | 340M |


# Training Dataset

The vocab and the model are pertrained on **100% of the PRD** scientific literature dataset.
 
The PRD dataset is provided by Public.Resource.Org, Inc. (“Public Resource”), 
a nonprofit organization based in California. This dataset was constructed from a corpus
of journal article files, from which We successfully extracted text from 75,496,055 articles from 178,928 journals.
The articles span across Arts & Humanities, Life Sciences & Biomedicine, Physical Sciences,
Social Sciences, and Technology. The distribution of articles is shown below.

![corpus pie chart](corpus_pie_chart.png)


# BibTeX entry and citation info
If using this model, please cite this paper:
```
@inproceedings{hong2023diminishing,
  title={The diminishing returns of masked language models to science},
  author={Hong, Zhi and Ajith, Aswathy and Pauloski, James and Duede, Eamon and Chard, Kyle and Foster, Ian},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2023},
  pages={1270--1283},
  year={2023}
}
```