X-ALMA-13B-Group1: Multilingual Translation and QA

X-ALMA-13B-Group1 is a 13 billion parameter model developed by Haoran Xu et al., extending the ALMA-R architecture to support an expanded set of 50 languages. This specific release focuses on Group 1 languages, which include English (en), Danish (da), Dutch (nl), German (de), Icelandic (is), Norwegian (no), Swedish (sv), and Afrikaans (af).

Key Capabilities

• Expanded Multilingual Support: Builds upon ALMA-R's foundation, significantly increasing language coverage.
• Plug-and-Play Architecture: Utilizes language-specific modules, allowing for flexible integration and targeted language support.
• Optimized Training: Incorporates a carefully designed training recipe for enhanced multilingual performance.
• Translation: Excels at translation tasks between supported languages.
• Multilingual Open-ended QA: Capable of performing question answering in the specified Group 1 languages.

Usage Recommendations

This model is provided as a merged model, where the language-specific module for Group 1 has been integrated into the base model. This is the recommended method for loading and using X-ALMA-13B-Group1 for translation and QA tasks in the supported languages. Alternatively, users can load the base model and then attach the language-specific LoRA module.