README.md

---
license: cc-by-sa-4.0
language:
- ko
tags:
- korean
---

# **KoBigBird-RoBERTa-large**

This is a large-sized Korean BigBird model introduced in our [paper](https://arxiv.org/abs/2309.10339).
The model draws heavily from the parameters of [klue/roberta-large](https://huggingface.co./klue/roberta-large) to ensure high performance.
By employing the BigBird architecture and incorporating the newly proposed TAPER, the language model accommodates even longer input lengths.

### How to Use

```python
from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("vaiv/kobigbird-roberta-large")
model = AutoModelForMaskedLM.from_pretrained("vaiv/kobigbird-roberta-large")
```

### Hyperparameters

![image/png](https://cdn-uploads.huggingface.co/production/uploads/62ce3886a9be5c195564fd71/bhuidw3bNQZbE2tzVcZw_.png)

### Results

Measurement on validation sets of the KLUE benchmark datasets

![image/png](https://cdn-uploads.huggingface.co/production/uploads/62ce3886a9be5c195564fd71/50jMYggkGVUM06n2v1Hxm.png)

### Limitations
While our model achieves great results even without additional pretraining, further pretraining can refine the positional representations more.

## Citation Information

```bibtex
@article{yang2023kobigbird,
    title={KoBigBird-large: Transformation of Transformer for Korean Language Understanding},
    author={Yang, Kisu and Jang, Yoonna and Lee, Taewoo and Seong, Jinwoo and Lee, Hyungjin and Jang, Hwanseok and Lim, Heuiseok},
    journal={arXiv preprint arXiv:2309.10339},
    year={2023}
}
```