SVILA: Slovenian Vision-Language Assistant
Overview
SVILA (Slovenian Vision-Language Assistant) is the first large-scale, general-purpose instruction-tuned VLM specifically for the Slovenian language. It addresses the digital divide by enabling advanced visual reasoning in a less-resourced language. The system is designed for diverse tasks including Visual Question Answering (VQA), visual grounding, and complex Optical Character Recognition (OCR). The models are based on the Google Gemma-3 series (4B and 12B parameters). They were fine-tuned using a custom Supervised Fine-Tuning (SFT) trainer on a curated dataset of over 1 million Slovenian text-image pairs. This dataset was constructed through a combination of automatic curation from Slovenian news media/Wikipedia and high-quality machine translation of the LLaVA-665k dataset.
Outcome
High-performance vision-language models capable of understanding and describing visual content in fluent Slovenian. The project provides a blueprint for adapting state-of-the-art multimodal models to other less-resourced languages and includes a manually curated test set for rigorous evaluation.
Service Card
| Category | AI products |
| Domain | Multimodal AI, Computer Vision, Slovenian NLP |
| Vertical | Digital Society |
| Target users | Tourism Industry, Media Houses, Slovenian Tech Companies Requiring Image-to-Text Capabilities |
| Service type | Multimodal Foundation Model (VLM) |
| Outputs | Fine-tuned weights for SVILA-4B and SVILA-12B models; access to the 1M+ curated Slovenian text-image pairs via CLARIN.SI; task support for image description, visual reasoning, and Slovenian-centric OCR. |
| Requirements | Inference: NVIDIA GPU (e.g., A100 or H100) for optimal performance. Software: Compatible with Hugging Face Transformers and VLM-specific libraries. |
| Limitations | Performance is primarily optimized for the Slovenian language; while based on multilingual Gemma-3, specific visual-cultural nuances are tailored to the Slovenian context. |
| Availability | Q4/2026 |
| Licensing | Models: MIT License; Dataset: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) |
| Visual content | assets/service-visuals/svila-slovenian-vision-language-assistant.png |
| Visual caption |