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. This is the card for the Gemma2 9B CPT SEA-LIONv3 base model which has undergone continued pre-training from the base Gemma-2-9B model.

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

Model Details

Model Description

The continued pre-training data for Gemma2 9B CPT SEA-LIONv3 base model encompasses approximately 200B tokens.

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

For tokenisation, the model employs the default tokenizer used in Gemma-2-9B. The model has a context length of 8192.

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/

Training Details

Data

Gemma2 9B CPT SEA-LIONv3 base model was continued pre-trained on 200B tokens of the following data:

Data Source	Unique Tokens (B)	Multiplier	Total Tokens (B)	Percentage (%)
StackV2	40.0	1	40.0	20.00
Wiki* + News* - English	5.0	1	5.0	2.50
Fineweb-Edu	7.5	1	7.5	3.75
Dolma Project Gutenberg	5.0	1	5.0	2.50
Dolma arXiv	1.7	1	1.7	0.83
Dolma StackExchange	1.7	1	1.7	0.83
Dolma Semantic Scholar	1.7	1	1.7	0.83
Dolma OpenWebMath	2.5	1	2.5	1.25
Dolma Algebraic Stack	2.5	1	2.5	1.25
Dolma Flan	5.0	1	5.0	2.50
Dolma Reddit	5.0	1	5.0	2.50
Dolma Megawika	5.0	1	5.0	2.50
Dolma CC News	7.5	1	7.5	3.75
Wiki* + News* - Chinese	3.5	4	14.0	7.00
SEA-LION Pile - Chinese	12.0	1	12.0	6.00
Wiki* + News* - Vietnamese	2.4	4	9.4	4.70
VinBigData - Vietnamese	2.1	4	8.2	4.10
SEA-LION Pile - Vietnamese	8.4	1	8.4	4.20
Wiki* + News* - Indonesian	1.3	4	5.2	2.60
SEA-LION Pile - Indonesian	20.8	1	20.8	10.40
Wiki* + News* + WangChanBERTa - Thai	1.3	4	5.2	2.60
SEA-LION Pile - Thai	14.8	1	14.8	7.40
Wiki* + News - Filipino	0.2	4	0.9	0.43
SEA-LION Pile - Filipino	2.1	1	2.1	1.07
Wiki* + News - Tamil	0.1	4	0.3	0.14
SEA-LION Pile - Tamil	0.7	1	0.7	0.36
Wiki* + News - Malay	0.1	4	0.6	0.29
SEA-LION Pile - Malay	1.4	1	1.4	0.71
Wiki* + News - Khmer	0.1	4	0.3	0.17
SEA-LION Pile - Khmer	2.3	1	2.3	1.13
Wiki* + News - Lao	0.0	4	0.1	0.03
SEA-LION Pile - Lao	0.3	1	0.3	0.17
Wiki* + News - Burmese	0.1	4	0.4	0.20
SEA-LION Pile - Burmese	2.6	1	2.6	1.30

Note:

All token counts are counted using Gemma2 tokenizer
Wiki* sources includes Wikipedia, Wiki Books, Wiki Source, Wiki Voyage and Fandom Wiki
News* sources includes VOA, Global Voices, MediaCorp, VinBigData-News
Tamil news is sourced with permission from Seithi

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+1 instances
Nvidia H100 80GB GPU	64+8
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

The Team

Chan Adwin, Choa Esther, Cheng Nicholas, Huang Yuli, 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, Teo Eng Sipp Leslie, Teo Wei Yi, Tjhi William, Teng Walter, 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 National Research Foundation, 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 base 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 claim, damages, or other liability 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}
}

aisingapore
/

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