SEA-LION-BERT

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

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

How To Use

from transformers import AutoModelForMaskedLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('aisingapore/sealion-bert-base', trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained('aisingapore/sealion-bert-base', 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 790B 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 14 days

Configuration

HyperParameter SEA-LION-BERT
Precision bfloat16
Optimizer decoupled_adamw
Scheduler linear_decay_with_warmup
Learning Rate 5e-4
Global Batch Size 448
Micro Batch Size 56

Technical Specifications

Model Architecture and Objective

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

Parameter SEA-LION-BERT
Layers 12
d_model 768
head_dim 12
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 Eng Sipp 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 using this SEA-LION Inquiry Form

Downloads last month
39
Safetensors
Model size
311M params
Tensor type
F32
·
Inference Examples
Inference API (serverless) has been turned off for this model.