SeaLLMs-v3 - Large Language Models for Southeast Asia
We introduce SeaLLMs-v3, the latest series of the SeaLLMs (Large Language Models for Southeast Asian languages) family. It achieves state-of-the-art performance among models with similar sizes, excelling across a diverse array of tasks such as world knowledge, mathematical reasoning, translation, and instruction following. In the meantime, it was specifically enhanced to be more trustworthy, exhibiting reduced hallucination and providing safe responses, particularly in queries closed related to Southeast Asian culture.
🔥 Highlights
- State-of-the-art performance compared to open-source models of similar sizes, evaluated across various dimensions such as human exam questions, instruction-following, mathematics, and translation.
- Significantly enhanced instruction-following capability, especially in multi-turn settings.
- Ensures safety in usage with significantly reduced instances of hallucination and sensitivity to local contexts.
Uses
SeaLLMs is tailored for handling a wide range of languages spoken in the SEA region, including English, Chinese, Indonesian, Vietnamese, Thai, Tagalog, Malay, Burmese, Khmer, Lao, Tamil, and Javanese.
This page introduces the SeaLLMs-v3-1.5B model, which can be easily fine-tuned for your specific downstream tasks, especially in SEA languages. Note that this is a base model, if you are looking for a model that can be directly applicable to your downstream applications, you may want to check the chat version model: SeaLLMs-v3-1.5B-Chat.
Evaluation
Evaluation
We evaluate SeaLLMs-v3-1.5B mainly using human exam questions.
Multilingual World Knowledge - M3Exam
M3Exam consists of local exam questions collected from each country. It reflects the model's world knowledge (e.g., with language or social science subjects) and reasoning abilities (e.g., with mathematics or natural science subjects).
| Model | en | zh | id | th | vi | avg | avg_sea |
|---|---|---|---|---|---|---|---|
| Gemma-2B | 0.411 | 0.267 | 0.296 | 0.283 | 0.313 | 0.314 | 0.297 |
| Sailor-1.8B | 0.270 | 0.239 | 0.250 | 0.261 | 0.260 | 0.256 | 0.257 |
| Sailor-4B | 0.387 | 0.295 | 0.275 | 0.296 | 0.311 | 0.313 | 0.294 |
| Qwen2-1.5B | 0.628 | 0.753 | 0.409 | 0.352 | 0.443 | 0.517 | 0.401 |
| SeaLLMs-v3-1.5B | 0.635 | 0.745 | 0.424 | 0.371 | 0.465 | 0.528 | 0.420 |
Multilingual World Knowledge - MMLU
MMLU questions are translated to SEA languages for evaluation, which primarily tests the cross-lingual alignment of the model as the required knowledge is still mainly Western-focused.
| Model | en | zh | id | th | vi | avg | avg_sea |
|---|---|---|---|---|---|---|---|
| Gemma-2B | 0.374 | 0.304 | 0.315 | 0.292 | 0.305 | 0.318 | 0.304 |
| Sailor-1.8B | 0.293 | 0.251 | 0.268 | 0.256 | 0.256 | 0.265 | 0.260 |
| Sailor-4B | 0.333 | 0.267 | 0.299 | 0.278 | 0.282 | 0.292 | 0.286 |
| Qwen2-1.5B | 0.552 | 0.491 | 0.426 | 0.366 | 0.398 | 0.447 | 0.397 |
| SeaLLMs-v3-1.5B | 0.553 | 0.487 | 0.443 | 0.377 | 0.423 | 0.456 | 0.414 |
Acknowledgement to Our Linguists
We would like to express our special thanks to our professional and native linguists, Tantong Champaiboon, Nguyen Ngoc Yen Nhi and Tara Devina Putri, who helped build, evaluate, and fact-check our sampled pretraining and SFT dataset as well as evaluating our models across different aspects, especially safety.
Citation
If you find our project useful, we hope you would kindly star our repo and cite our work as follows:
@article{damonlp2024seallm3,
author = {Wenxuan Zhang*, Hou Pong Chan*, Yiran Zhao*, Mahani Aljunied*,
Jianyu Wang*, Chaoqun Liu, Yue Deng, Zhiqiang Hu, Weiwen Xu,
Yew Ken Chia, Xin Li, Lidong Bing},
title = {SeaLLMs 3: Open Foundation and Chat Multilingual Large Language Models for Southeast Asian Languages},
year = {2024},
url = {https://arxiv.org/abs/2407.19672}
}
Corresponding Author: l.bing@alibaba-inc.com