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
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 | linear_decay_with_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
- Downloads last month
- 26