🚀 Llama-SEA-LION-v3-8B
SEA-LION is a collection of Large Language Models (LLMs) pre - trained and instruct - tuned for the Southeast Asia (SEA) region. Llama-SEA-LION-v3-8B is a multilingual model pre - trained on about 200B tokens across 11 SEA languages.
🚀 Quick Start
This section provides a high - level overview of the Llama-SEA-LION-v3-8B model. For more detailed usage, please refer to the corresponding sections below.
✨ Features
- Multilingual Capability: Llama-SEA-LION-v3-8B has been continued pre - trained on approximately 200B tokens across 11 Southeast Asian languages, including Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai and Vietnamese.
- Based on Llama 3.1 Architecture: It is developed based on the Llama 3.1 architecture, specifically the Llama-3.1-8B-Instruct decoder model.
- Evaluated on Multiple Benchmarks: The model has been evaluated on general language capabilities and constraint - following behaviour using benchmarks like SEA - HELM and SEA - IFEval.
📚 Documentation
Model Details
Model Description
We created Llama-SEA-LION-v3-8B by performing continued pre - training in English and SEA languages on Llama-3.1-8B-Instruct, a decoder model using the Llama 3.1 architecture. The model uses the default tokenizer of Llama 3.1 8B Instruct for tokenization.
Benchmark Performance
We evaluated Llama-SEA-LION-v3-8B on general language capabilities and constraint - following behaviour.
- General Language Capabilities and Constraint - following Behaviour: We used the SEA - HELM evaluation benchmark for evaluating general language capabilities across various tasks such as Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal) and Natural Language Inference (NLI). The evaluation was done five - shot with native prompts on a sample of 100 - 1000 instances for each dataset.
- SEA - IFEval: Based on IFEval, we implemented SEA - IFEval to compare the model's ability to follow specific constraints in English and SEA languages. Linguists and native speakers in the team filtered, localized and translated the datasets to ensure the examples were reasonable, meaningful and natural.
For more details on Llama-SEA-LION-v3-8B benchmark performance, please refer to the SEA - HELM leaderboard, https://leaderboard.sea - lion.ai/.
Technical Specifications
Infrastructure
Llama-SEA-LION-v3-8B was trained using MosaicML Composer on the following hardware:
Training Details |
Llama-SEA-LION-v3-8B |
AWS p5e.48xlarge |
8 instances |
Nvidia H200 140GB GPU |
64 |
Training Duration |
136 Hours |
Configuration
HyperParameter |
Llama-SEA-LION-v3-8B |
Precision |
bfloat16 |
Optimizer |
decoupled_adamw |
Scheduler |
weight_stable_decay |
Learning Rate |
1.0e-5 |
Global Batch Size |
512 |
Data
Llama-SEA-LION-v3-8B was continued pre - trained on 200B tokens of the following data:
Language |
Source |
Total Tokens (B) |
Percentage (%) |
Total percentage (%) |
Code |
StackV2 |
40 |
20 |
20 |
English |
Dolma |
37.5 |
18.75 |
25 |
|
Fineweb-Edu |
7.5 |
3.75 |
|
|
Others |
5 |
2.5 |
|
Chinese |
SEA-LION Pile v1 |
12 |
6 |
13 |
|
Others |
14 |
7 |
|
Vietnamese |
SEA-LION Pile v1 |
8.4 |
4.2 |
13 |
|
VinBigData |
16 |
8 |
|
|
Others |
1.6 |
0.8 |
|
Indonesian |
SEA-LION Pile v1 |
7 |
3.5 |
13 |
|
SEA-LION Pile v2 |
7 |
3.5 |
|
|
Others |
12 |
6 |
|
Thai |
SEA-LION Pile v1 |
10.7 |
5.35 |
10 |
|
WangChanBERTa |
8.5 |
4.25 |
|
|
Others |
0.8 |
0.4 |
|
Filipino - Malay - Tamil |
SEA-LION Pile v1, AI4Bharat Sangraha |
4.28 |
2.14 |
3 |
|
Others |
1.72 |
0.86 |
|
Khmer - Lao - Burmese |
SEA-LION Pile v1 |
5.2 |
2.6 |
3 |
|
Others |
0.8 |
0.4 |
|
Note:
- All token counts are counted using Llama 3.1 8B Instruct tokenizer.
- SEA - LION Pile v1 is processed from Common Crawl WET, which is published [here](https://huggingface.co/datasets/aisingapore/sea - lion - pile). The cutoff date of this version is September 2020.
- SEA - LION Pile v2 is processed from Common Crawl WARC from October 2020 to April 2024.
- Tamil data from Sangraha is published here. The paper can be found here.
- Tamil news is sourced with permission from Seithi
Call for Contributions
We encourage researchers, developers, and language enthusiasts to actively contribute to the enhancement and expansion of SEA - LION. Contributions can involve identifying and reporting bugs, sharing pre - training, instruction, and preference data, improving documentation usability, proposing and implementing new model evaluation tasks and metrics, or training versions of the model in additional Southeast Asian languages. Join us in shaping the future of SEA - LION by sharing your expertise and insights to make these models more accessible, accurate, and versatile. Please check out our GitHub for further information on the call for contributions.
The Team
Chan Adwin, Cheng Nicholas, Choa Esther, Huang Yuli, Hulagadri Adithya Venkatadri, Lau Wayne, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Limkonchotiwat Peerat, Liu Bing Jie Darius, Montalan Jann Railey, Ng Boon Cheong Raymond, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Brandon, Ong Tat - Wee David, Ong Zhi Hao, Rengarajan Hamsawardhini, Siow Bryan, Susanto Yosephine, Tai Ngee Chia, Tan Choon Meng, Teng Walter, Teo Eng Sipp Leslie, Teo Wei Yi, Tjhi William, Yeo Yeow Tong, Yong Xianbin
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 the National Research Foundation or the National University of Singapore.
Contact
For more info, please contact us using this SEA - LION Inquiry Form.
Link to SEA - LION's GitHub repository.
Disclaimer
This is the repository for the commercial instruction - tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine - tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.
References
Thai Pre - Training Data Reference
@misc{lowphansirikul2021wangchanberta,
title={WangchanBERTa: Pretraining transformer-based Thai Language Models},
author={Lalita Lowphansirikul and Charin Polpanumas and Nawat Jantrakulchai and Sarana Nutanong},
year={2021},
eprint={2101.09635},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
📄 License
The model is released under the [Llama 3.1 Community License](https://github.com/meta - llama/llama - models/blob/main/models/llama3_1/LICENSE).
Property |
Details |
Model Type |
Decoder |
Training Data |
Approximately 200B tokens across 11 SEA languages: Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai and Vietnamese. |
Developed by |
Products Pillar, AI Singapore |
Funded by |
Singapore NRF |
Languages supported |
Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai, Vietnamese |
License |
[Llama 3.1 Community License](https://github.com/meta - llama/llama - models/blob/main/models/llama3_1/LICENSE) |