haoranxu/X-ALMA-13B-Group3

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Aug 23, 2024License:mitArchitecture:Transformer0.0K Open Weights Cold

haoranxu/X-ALMA-13B-Group3 is a 13 billion parameter multilingual language model developed by Haoran Xu et al., building on the ALMA-R architecture. It features a plug-and-play design with language-specific modules, supporting 50 languages, with this specific release optimized for English, Bulgarian, Macedonian, Serbian, Ukrainian, and Russian. The model is designed for quality translation at scale and multilingual open-ended QA, utilizing a carefully designed training recipe and a 4096-token context length.

Loading preview...

X-ALMA-13B-Group3 Overview

X-ALMA-13B-Group3 is a 13 billion parameter multilingual language model developed by Haoran Xu et al., part of the broader X-ALMA family. It extends the ALMA-R architecture to support 50 languages through a novel plug-and-play module design and an adaptive rejection mechanism for high-quality translation. This specific model release focuses on a Group 3 set of languages: English (en), Bulgarian (bg), Macedonian (mk), Serbian (sr), Ukrainian (uk), and Russian (ru).

Key Capabilities

  • Multilingual Translation: Optimized for translation across its supported Group 3 languages.
  • Plug-and-Play Architecture: Utilizes language-specific LoRA modules that can be loaded with a base model or as a pre-merged model.
  • Multilingual Open-Ended QA: Capable of performing question answering tasks in the supported languages.
  • Scalable Design: The X-ALMA framework is designed for efficient scaling to a wide range of languages.

Usage and Integration

The model offers flexible loading options:

  • Recommended: Load the pre-merged model, where the language-specific module is already integrated.
  • Alternative: Load the haoranxu/X-ALMA-13B-Pretrain base model and then attach the X-ALMA-13B-Group3 LoRA module.
  • Advanced: Load the full haoranxu/X-ALMA model with all language-specific modules, requiring substantial GPU memory.

Good for

  • Developers requiring high-quality machine translation between English and the specified Eastern European languages.
  • Applications needing multilingual question answering capabilities in these language pairs.
  • Research into modular, scalable multilingual LLM architectures.