LLaMAX/LLaMAX3-8B-Alpaca

LLaMAX3-8B-Alpaca: Multilingual Translation and Instruction Following

LLaMAX3-8B-Alpaca is an 8 billion parameter language model from the LLaMAX series, engineered to provide robust multilingual translation capabilities alongside strong instruction-following. This model was developed by collecting extensive training sets in 102 languages for continued pre-training of Llama-2 and then fine-tuned using the Alpaca English instruction dataset.

Key Capabilities

• Extensive Multilingual Translation: Supports translation between more than 100 languages, outperforming similarly scaled LLMs in this domain.
• Enhanced Translation Performance: Achieves an average spBLEU score improvement of over 5 points compared to LLaMA3-8B-Alpaca on the Flores-101 dataset, demonstrating superior performance across various language pairs (e.g., en-X, X-en, zh-X, X-zh).
• Instruction Following: Retains strong instruction-following capabilities, making it suitable for tasks beyond just translation.

Good For

• Multilingual Applications: Ideal for developers building applications requiring accurate translation across a wide array of languages.
• Cross-Lingual Communication: Facilitating communication and content generation in diverse linguistic contexts.
• Research in Multilingual LLMs: Serving as a robust foundation model for further research into multilingual natural language processing.

For more technical details, refer to the LLaMAX paper.