Infomaniak-AI/vllm-translategemma-12b-it

Overview

Infomaniak-AI/vllm-translategemma-12b-it is a 12 billion parameter instruction-tuned model derived from Google's TranslateGemma family, designed for efficient translation across 55 languages. This specific version is optimized for deployment with vLLM, featuring critical modifications for compatibility and performance.

Key Modifications for vLLM

• Chat Template: The original structured payload with source_lang_code and target_lang_code fields is replaced with a delimiter-based format within the message content, enabling direct language code encoding compatible with vLLM.
  • Format: <<<source>>>{source_lang}<<<target>>>{target_lang}<<<text>>>{text_to_translate}
• Model Configuration (RoPE): The RoPE parameters have been simplified from a dual full_attention and sliding_attention structure to a single factor and rope_type for vLLM compatibility.
• Generation Configuration: The End-Of-Sequence (EOS) token has been corrected from <end_of_turn> to <eos> for proper sequence termination.

Core Capabilities

• Multilingual Translation: Handles translation tasks across 55 languages.
• Multimodal Input: Supports both text-to-text translation and text extraction-and-translation from images (normalized to 896x896 resolution).
• Resource-Efficient: Its relatively small size (12B parameters) allows deployment in environments with limited resources, such as laptops, desktops, or private cloud infrastructure.

Intended Usage

This model is explicitly designed for producing text translations from either textual or image inputs. It offers superior performance compared to other comparably-sized open models for translation tasks, as indicated by benchmark results like Comet and MetricX scores on WMT24++ and WMT25 datasets. The model was trained using JAX and ML Pathways on Google's TPUs, emphasizing performance and scalability.