AI Singapore has launched SEA-LION v2, the latest addition to its open-source language model series aimed at understanding and reflecting the linguistic and cultural diversity of Southeast Asia.
Dr. Leslie Teo, Senior Director at AI Singapore and the project lead, stated on LinkedIn that this updated model is designed to offer more precise and contextually relevant language processing abilities tailored to the specific requirements of the region.
SEA-LION v2 is aligned with AI Singapore’s objectives of enhancing AI capabilities, generating social and economic impact, nurturing local talent, fostering an AI ecosystem, and positioning Singapore as a global leader in AI. The model utilizes advanced open-source frameworks with ongoing pre-training and fine-tuning for Southeast Asia.
Unlike its predecessor, which was trained from scratch, SEA-LION v2 is based on Meta’s Llama 3, leveraging 64x Nvidia H100 GPUs for faster training and performance improvement.
The model was part of the National Multimodal LLM Programme, supported by a SG$70 million government fund to develop AI talent. SEA-LION v2 has shown enhanced performance in regional languages while retaining general capabilities.
- Continued pre-training and fine-tuning based on Llama 3 architecture
- Multilingual capabilities in English, Bahasa Indonesia, Thai, Vietnamese, and Tamil
- Trained with around 50 billion tokens from Southeast Asian languages
- Licensed under the Meta Llama 3 Community License
SEA-LION v2 is now available for download on HuggingFace in different models. Although no online demo is available, the instruction-tuned model supports basic interactions when deployed correctly.
Future plans include further collaborations with Google’s Gemma 2 and AI startup Reka for model enhancements.
Fundraising or getting ready to raise capital for your startup? Connect with over 400 SEA investors on e27 Pro for insights and networking opportunities.
Image credit: Canva Pro
The post AI Singapore releases SEA-LION v2 designed to understand SEA’s linguistic, cultural diversity appeared first on e27.