Llama-SEA-LION-v3-70B is a multilingual model which has undergone continued pre-training on approximately 200B tokens across 11 SEA languages: Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai and Vietnamese.

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: Llama 3.1 Community License

Model Details

Model Description

We performed continued pre-training in English and SEA languages on Llama-3.1-70B-Instruct, a decoder model using the Llama 3.1 architecture, to create Llama-SEA-LION-v3-70B.

For tokenisation, the model employs the default tokenizer used in Llama 3.1 70B Instruct.

Benchmark Performance

We evaluated Llama-SEA-LION-v3-70B on general language capabilities and constraint-following behaviour.

General Language Capabilities and Constraint-following Behaviour

For the evaluation of general language capabilities, we employed the SEA-HELM 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 Summarisation (Abssum), 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.

Following the implementation of IFEval in OpenLLM leaderboard, we also implement SEA-IFEval to provide a comparison of the ability of the model to follow specific constraints in English and in SEA languages.

SEA-IFEval

Based on IFEval, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.

SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).

For more details on Llama-SEA-LION-v3-70B benchmark performance, please refer to the SEA-HELM leaderboard, https://leaderboard.sea-lion.ai/.

Technical Specifications

Infrastructure

Llama-SEA-LION-v3-70B was trained in two stages using MosaicML Composer on the following hardware:

Stage	Training Details	Llama-SEA-LION-v3-70B
First Stage	AWS p5e.48xlarge	8 instances
	Nvidia H200 140GB GPU	64
	Training Duration	200 hrs (step 0 - 9000)
Second Stage	SingTel HGX-100	16 instances
	Nvidia H100 80GB GPU	128
	Training Duration	495 hrs (step 9000 - 47684)

Configuration

HyperParameter	Llama-SEA-LION-v3-70B
Precision	bfloat16
Optimizer	decoupled_adamw
Scheduler	weight_stable_decay
Learning Rate	1.0e-5
Global Batch Size	512

Data

Llama-SEA-LION-v3-70B 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 70B Instruct 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 data from Sangraha is published here. The paper can be found here.
Tamil news is sourced with permission from Seithi

Lineage & Versioning

The training counts and dataset mixture details reported in this model card reflect the exact constructed training pool consumed during this specific model run. Figures may differ slightly from public dataset releases, which represent downloadable open-source subsets of the broader corpus.

This model card serves as an immutable record of the final released checkpoint. Consequently, the hardware specifications, compute hours, and precise dataset volumes (such as final filtered instruction counts) reported here reflect the exact production run used to generate this specific artifact. These figures may differ from the aggregate totals, pre-filtered data pools, or preliminary experimental runs (e.g., initial H100 benchmarks) documented in our accompanying research papers.

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

bibtex

@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/Llama-SEA-LION-v3-70B

Llama-SEA-LION-v3-70B