aisingapore/Gemma-SEA-LION-v3-9B-IT

Gemma-SEA-LION-v3-9B-IT Overview

Gemma-SEA-LION-v3-9B-IT is an instruction-tuned large language model developed by AI Singapore, part of the SEA-LION (Southeast Asian Languages In One Network) collection. This model is based on the Gemma2 architecture and has been specifically fine-tuned for both English and a range of Southeast Asian languages, including Burmese, Chinese, Filipino, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tamil, Thai, and Vietnamese. It utilizes the default Gemma-2-9B tokenizer and supports a context length of 8192 tokens.

Key Capabilities and Features

• Multilingual Proficiency: Instruction-tuned for a broad spectrum of Southeast Asian languages, enhancing performance in regional contexts.
• Instruction Following: Evaluated on datasets like SEA-IFEval and SEA-MTBench, demonstrating its ability to adhere to prompt constraints and engage in multi-turn conversations.
• General Language Understanding: Performance assessed using the SEA-HELM benchmark across tasks such as Question Answering, Sentiment Analysis, Translation, and Summarization.
• Decoder Architecture: Built upon the Gemma2 decoder model for robust language generation.

Important Considerations

• No System Prompt/Tool Calling: The model has not been trained to utilize system prompts or tool-calling functionalities.
• Safety Alignment: This model has not been aligned for safety, and users are advised to perform their own safety fine-tuning and implement security measures.
• Limitations: Like other LLMs, it may exhibit hallucinations and occasional inconsistencies in reasoning.