aisingapore/gemma2-9b-cpt-sea-lionv3-base

Gemma2 9B CPT SEA-LIONv3

SEA-LION is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for the Southeast Asia (SEA) region.

Gemma2 9B CPT SEA-LIONv3 Base is a multilingual model which has undergone continued pre-training on approximately 200B tokens across the 11 official Southeast Asian languages: English, Chinese, Vietnamese, Indonesian, Thai, Tamil, Filipino, Malay, Khmer, Lao, Burmese.

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

• Developed by: Products Pillar, AI Singapore
• Funded by: Singapore NRF
• Model type: Decoder
• Languages supported: Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai, Vietnamese
• License: Gemma Community License

Model Details

Model Description

We performed continued pre-training in English and ASEAN languages on Gemma-2-9B, a decoder model using the Gemma 2 architecture, to create Gemma2 9B CPT SEA-LIONv3 Base.

For tokenisation, the model employs the default tokenizer used in Gemma 2 9B.

Benchmark Performance

We evaluated Gemma2 9B CPT SEA-LIONv3 base model on general language capabilities.

General Language Capabilities

For the evaluation of general language capabilities, we employed the SEA HELM (also known as BHASA) evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarization (Summ), Causal Reasoning (Causal) and Natural Language Inference (NLI).

Note: SEA HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.

The evaluation was done five-shot with native prompts on a sample of 100-1000 instances for each dataset.

For more details on Gemma2 9B CPT SEA-LIONv3 base benchmark performance, please refer to the SEA HELM leaderboard, https://leaderboard.sea-lion.ai/

Technical Specifications

Infrastructure

Gemma2 9B CPT SEA-LIONv3 was trained using MosaicML Composer on the following hardware:

Training Details	Gemma2 9B CPT SEA-LIONv3
SingTel HGX-100	8 instances
Nvidia H100 80GB GPU	64
Training Duration	10 days

Configuration

HyperParameter	Gemma2 9B CPT SEA-LIONv3
Precision	bfloat16
Optimizer	decoupled_adamw
Scheduler	weight_stable_decay
Learning Rate	1.0e-5
Global Batch Size	512
Micro Batch Size	1

Data

Gemma2 9B CPT SEA-LIONv3 base model 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	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 Gemma 2 9B tokenizer
• SEA-LION Pile v1 is processed from Common Crawl WET, which is published here. 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 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}
}