LazarusNLP/NusaBERT-large

NusaBERT Large

NusaBERT Large is a multilingual encoder-based language model based on the BERT architecture. We conducted continued pre-training on open-source corpora of sabilmakbar/indo_wiki, acul3/KoPI-NLLB, and uonlp/CulturaX. On a held-out subset of the corpus, our model achieved:

• eval_accuracy: 0.7117
• eval_loss: 1.3268
• perplexity: 3.7690

This model was trained using the 🤗Transformers PyTorch framework. All training was done on an NVIDIA H100 GPU. LazarusNLP/NusaBERT-large is released under Apache 2.0 license.

Model Detail

• Developed by: LazarusNLP
• Finetuned from: IndoBERT Large p1
• Model type: Encoder-based BERT language model
• Language(s): Indonesian, Acehnese, Balinese, Banjarese, Buginese, Gorontalo, Javanese, Banyumasan, Minangkabau, Malay, Nias, Sundanese, Tetum
• License: Apache 2.0
• Contact: LazarusNLP

Use in 🤗Transformers

from transformers import AutoTokenizer, AutoModelForMaskedLM

model_checkpoint = "LazarusNLP/NusaBERT-large"

tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
model = AutoModelForMaskedLM.from_pretrained(model_checkpoint)

Training Datasets

Around 16B tokens from the following corpora were used during pre-training.

• Indonesian Wikipedia Data Repository
• KoPI-NLLB (Korpus Perayapan Indonesia)
• Cleaned, Enormous, and Public: The Multilingual Fuel to Democratize Large Language Models for 167 Languages

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 256
• eval_batch_size: 256
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 24000
• training_steps: 500000

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.0+cu118
• Datasets 2.17.1
• Tokenizers 0.15.2

Credits

NusaBERT Large is developed with love by:

Citation

@misc{wongso2024nusabert,
  title={NusaBERT: Teaching IndoBERT to be Multilingual and Multicultural}, 
  author={Wilson Wongso and David Samuel Setiawan and Steven Limcorn and Ananto Joyoadikusumo},
  year={2024},
  eprint={2403.01817},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}