Hugging Face Overview
SeaLLMs-v3-7B-Chat: Multilingual LLM for Southeast Asia
SeaLLMs-v3-7B-Chat is the latest 7.6 billion parameter model from the SeaLLMs family, specifically designed and fine-tuned for chat and instruction-following in Southeast Asian languages. It demonstrates strong performance across various benchmarks, outperforming models of similar sizes in key areas.
Key Capabilities & Differentiators
- Multilingual Proficiency: Optimized for a broad spectrum of SEA languages including English, Chinese, Indonesian, Vietnamese, Thai, Tagalog, Malay, Burmese, Khmer, Lao, Tamil, and Javanese.
- Enhanced Instruction Following: Significantly improved capability in understanding and executing multi-turn human instructions.
- Reduced Hallucination: Exhibits lower instances of hallucination and provides more contextually appropriate responses, particularly for Southeast Asian cultural queries.
- Strong Performance: Achieves competitive results in multilingual world knowledge (M3Exam), instruction-following (SeaBench), and multilingual mathematics (MGSM), with a notable 73.1% average on MGSM and 36.52 chrF score for translation.
- Improved Trustworthiness: Demonstrates high refusal F1 scores (e.g., 71.59% average) for non-existing entities and strong safety rates (e.g., 79.75% average) against harmful prompts across multiple languages.
Ideal Use Cases
This model is particularly well-suited for applications requiring robust multilingual communication and instruction-following capabilities within the Southeast Asian linguistic landscape, where cultural sensitivity and factual accuracy are paramount.