---
license: mit
language:
- en
- zh
- id
- ms
- th
- vi
- fil
- ta
- my
- km
- lo
---
# Note: This model has been deprecated. Please use our latest release at [gemma2-9b-cpt-sea-lionv3-base](https://huggingface.co/aisingapore/gemma2-9b-cpt-sea-lionv3-base).
<hr>

# SEA-LION

SEA-LION is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
The size of the models range from 3 billion to 7 billion parameters.
This is the card for the SEA-LION 7B base model.

SEA-LION stands for <i>Southeast Asian Languages In One Network</i>.


## Model Details

### Model Description

The SEA-LION model is a significant leap forward in the field of Natural Language Processing,
specifically trained to understand the SEA regional context.

SEA-LION is built on the robust MPT 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 encompasses 980B tokens.

- **Developed by:** Products Pillar, AI Singapore
- **Funded by:** Singapore NRF
- **Model type:** Decoder
- **Languages:** English, Chinese, Indonesian, Malay, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, Lao
- **License:** MIT License

### Performance Benchmarks

SEA-LION has an average performance on general tasks in English (as measured by Hugging Face's LLM Leaderboard):

| Model       | ARC   | HellaSwag | MMLU  | TruthfulQA | Average |
|-------------|:-----:|:---------:|:-----:|:----------:|:-------:|
| SEA-LION 7B | 39.93 | 68.51     | 26.87 |      35.09 | 42.60   |

## Training Details

### Data

SEA-LION was trained on 980B tokens of the following data:

| Data Source               | Unique Tokens | Multiplier | Total Tokens | Percentage |
|---------------------------|:-------------:|:----------:|:------------:|:----------:|
| RefinedWeb - English      |        571.3B |          1 |       571.3B |     58.20% |
| mC4 - Chinese             |         91.2B |          1 |        91.2B |      9.29% |
| mC4 - Indonesian          |         3.68B |          4 |        14.7B |      1.50% |
| mC4 - Malay               |         0.72B |          4 |         2.9B |      0.29% |
| mC4 - Filipino            |         1.32B |          4 |         5.3B |      0.54% |
| mC4 - Burmese             |          1.2B |          4 |         4.9B |      0.49% |
| mC4 - Vietnamese          |         63.4B |          1 |        63.4B |      6.46% |
| mC4 - Thai                |          5.8B |          2 |        11.6B |      1.18% |
| WangChanBERTa - Thai      |            5B |          2 |          10B |      1.02% |
| mC4 - Lao                 |         0.27B |          4 |         1.1B |      0.12% |
| mC4 - Khmer               |         0.97B |          4 |         3.9B |      0.40% |
| mC4 - Tamil               |         2.55B |          4 |        10.2B |      1.04% |
| the Stack - Python        |         20.9B |          2 |        41.8B |      4.26% |
| the Stack - Javascript    |         55.6B |          1 |        55.6B |      5.66% |
| the Stack - Shell         |         1.2B5 |          2 |         2.5B |      0.26% |
| the Stack - SQL           |         6.4B  |          2 |        12.8B |      1.31% |
| the Stack - Markdown      |         26.6B |          1 |        26.6B |      2.71% |
| RedPajama - StackExchange |         21.2B |          1 |        21.2B |      2.16% |
| RedPajama - ArXiv         |         30.6B |          1 |        30.6B |      3.12% |

### Infrastructure

SEA-LION was trained using [MosaicML Composer](https://github.com/mosaicml/composer)
on the following hardware:

| Training Details     | SEA-LION 7B  |
|----------------------|:------------:|
| AWS EC2 p4d.24xlarge | 32 instances |
| Nvidia A100 40GB GPU | 256          |
| Training Duration    | 22 days      |


### Configuration

| HyperParameter    | SEA-LION 7B        |
|-------------------|:------------------:|
| Precision         | bfloat16           |
| Optimizer         | decoupled_adamw    |
| Scheduler         | cosine_with_warmup |
| Learning Rate     | 6.0e-5             |
| Global Batch Size | 2048               |
| Micro Batch Size  | 4                  |


## Technical Specifications

### Model Architecture and Objective

SEA-LION is a decoder model using the MPT architecture.

| Parameter       | SEA-LION 7B |
|-----------------|:-----------:|
| Layers          | 32          |
| d_model         | 4096        |
| head_dim        | 32          |
| Vocabulary      | 256000      |
| Sequence Length | 2048        |


### Tokenizer Details

We sample 20M lines from the training data to train the tokenizer.<br>
The framework for training is [SentencePiece](https://github.com/google/sentencepiece).<br>
The tokenizer type is Byte-Pair Encoding (BPE).



## The Team

Lam Wen Zhi Clarence<br>
Leong Wei Qi<br>
Li Yier<br>
Liu Bing Jie Darius<br>
Lovenia Holy<br>
Montalan Jann Railey<br>
Ng Boon Cheong Raymond<br>
Ngui Jian Gang<br>
Nguyen Thanh Ngan<br>
Ong Tat-Wee David<br>
Rengarajan Hamsawardhini<br>
Susanto Yosephine<br>
Tai Ngee Chia<br>
Tan Choon Meng<br>
Teo Jin Howe<br>
Teo Eng Sipp Leslie<br>
Teo Wei Yi<br>
Tjhi William<br>
Yeo Yeow Tong<br>
Yong Xianbin<br>

## 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](https://forms.gle/sLCUVb95wmGf43hi6)

[Link to SEA-LION's GitHub repository](https://github.com/aisingapore/sealion)


## Disclaimer

This 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
```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}
}
```