INSAIT-Institute/MamayLM-Gemma-3-4B-IT-v1.0

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:4.3BQuant:BF16Context Size:32kPublished:Aug 29, 2025License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Warm

INSAIT-Institute/MamayLM-Gemma-3-4B-IT-v1.0 is a 4.3 billion parameter instruction-tuned causal language model developed by INSAIT, based on Google's Gemma 3 4B models. It was continuously pre-trained and fine-tuned to achieve outstanding Ukrainian cultural and linguistic capabilities while retaining strong English performance. This model excels in logical reasoning, mathematics, and knowledge tasks across both languages, outperforming similarly sized models in Ukrainian benchmarks. It is designed for applications requiring robust multilingual understanding, particularly in Ukrainian and English.

Loading preview...

Overview

INSAIT-Institute/MamayLM-Gemma-3-4B-IT-v1.0 is a 4.3 billion parameter language model developed by INSAIT, built upon Google's Gemma 3 4B base and instruction-tuned models. This model has undergone continuous pre-training on a large, filtered dataset combining data mixing and model merging techniques. This process has endowed MamayLM with exceptional Ukrainian linguistic and cultural understanding, while also preserving its strong English performance.

Key Capabilities

  • Bilingual Proficiency: Demonstrates outstanding performance in Ukrainian and maintains excellent English capabilities.
  • Broad Task Performance: Evaluated across various benchmarks including Winogrande, Hellaswag, ARC, TriviaQA, GSM-8k, MMLU, and IFEval, showcasing strong logical reasoning, mathematical, and general knowledge skills.
  • Ukrainian Specific Knowledge: Excels in Ukrainian-specific benchmarks like ZNO (Ukrainian high school curriculum) and ZNO-Vision, indicating deep cultural and educational context understanding.
  • Instruction Following: Instruction-fine-tuned on a newly constructed Ukrainian dataset, enhancing its ability to follow complex instructions.

Use Cases

MamayLM-Gemma-3-4B-IT-v1.0 is ideal for applications requiring high-quality language generation and understanding in both Ukrainian and English. Its specialized training makes it particularly suitable for:

  • Ukrainian-centric content creation and analysis.
  • Educational tools and platforms for Ukrainian curricula.
  • Multilingual chatbots and virtual assistants operating in Ukrainian and English.
  • Research and development in Ukrainian natural language processing.