Modelpublic

aisingapore/sealion-bert-large

sourceHugging Facemitupdated 3y agoView on Hugging Face

6likes41downloads

Model Card

SEA-LION-BERT

SEA-LION stands for Southeast Asian Languages In One Network.

This is the card for the SEA-LION-BERT large model.

How To Use

python

from transformers import AutoModelForMaskedLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('aisingapore/sealion-bert-large', trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained('aisingapore/sealion-bert-large', trust_remote_code=True)

# prepare input
text = "Give me a <|mask|>!!!"
encoded_input = tokenizer(text, return_tensors='pt')

Model Details

Model Description

The SEA-LION-BERT model is built on the MosaicBERT architecture and has a vocabulary size of 256K.

For tokenization, the model employs our custom SEABPETokenizer, which is specially tailored for SEA languages, ensuring optimal model performance.

The training data for SEA-LION-BERT encompasses 980B tokens.

—Developed by: Products Pillar, AI Singapore
—Funded by: Singapore NRF
—Model type: Encoder
—Languages: English, Chinese, Indonesian, Malay, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, Lao
—License: MIT License

Training Details

Data

SEA-LION was trained on 790B tokens of the following data:

Data Source	Tokens	Percentage
RefinedWeb - English	571.3B	72.26%
mC4 - Chinese	91.2B	11.54%
mC4 - Indonesian	14.7B	1.86%
mC4 - Malay	2.9B	0.36%
mC4 - Filipino	5.3B	0.67%
mC4 - Burmese	4.9B	0.61%
mC4 - Vietnamese	63.4B	8.02%
mC4 - Thai	21.6B	2.74%
mC4 - Lao	1.1B	0.14%
mC4 - Khmer	3.9B	0.50%
mC4 - Tamil	10.2B	1.29%

Infrastructure

SEA-LION was trained using MosaicML Composer on the following hardware:

Training Details	SEA-LION-BERT
Nvidia A100 40GB GPU	4
Training Duration	22 days

Configuration

HyperParameter	SEA-LION-BERT
Precision	bfloat16
Optimizer	decoupled_adamw
Scheduler	lineardecaywith_warmup
Learning Rate	1e-4
Global Batch Size	720
Micro Batch Size	36

Technical Specifications

Model Architecture and Objective

SEA-LION-BERT is an encoder model using the MosaicBERT architecture.

Parameter	SEA-LION-BERT
Layers	24
d_model	1024
head_dim	16
Vocabulary	256000
Sequence Length	128

Tokenizer Details

We sample 20M lines from the training data to train the tokenizer. The framework for training is SentencePiece. The tokenizer type is Byte-Pair Encoding (BPE).

The Team

Montalan Jann Railey Nguyen Thanh Ngan Rengarajan Hamsawardhini Teo Leslie Tjhi William

Acknowledgements

AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.

Contact

For more info, please contact us at sealion@aisingapore.org