aisingapore/Llama-SEA-LION-v3.5-70B-R

Llama-SEA-LION-v3.5-70B-R: A Hybrid LLM for Southeast Asian Languages

Llama-SEA-LION-v3.5-70B-R is a 70-billion parameter decoder-only model developed by AI Singapore, building upon the Llama 3.1 architecture. It is part of the SEA-LION (Southeast Asian Languages In One Network) collection, specifically designed and optimized for the Southeast Asian region. This model features a substantial 128k context length and supports a wide array of languages including Burmese, Chinese, English, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese.

Key Capabilities and Features

• Multilingual Proficiency: Instruction-tuned in English and several SEA languages (Filipino, Indonesian, Tamil, Thai, Vietnamese) for enhanced regional applicability.
• Hybrid Functionality: Capable of handling both complex reasoning tasks and general text generation, with mode selection managed via the tokenizer's chat template.
• Thinking Mode Toggle: Users can switch between a default 'thinking_mode="on"' for reasoning and 'thinking_mode="off"' for standard generations, offering flexible control over model behavior.
• Comprehensive Evaluation: Benchmarked using SEA-HELM for general language capabilities (QA, Sentiment, Translation, Summarization, Reasoning) and SEA-IFEval/SEA-MTBench for instruction-following, with results available on the SEA-HELM leaderboard.

Use Cases and Considerations

This model is ideal for applications requiring strong performance in Southeast Asian languages, particularly for tasks involving instruction-following, multi-turn conversations, and complex reasoning. Developers should note that the model has not been aligned for safety, and users are advised to implement their own safety fine-tuning and validation measures. Like other LLMs, it may exhibit limitations such as hallucination and occasional inconsistencies in reasoning.