lapa-llm/lapa-v0.1.3-instruct
lapa-llm/lapa-v0.1.3-instruct is a 12 billion parameter instruction-tuned large language model developed by a team of Ukrainian researchers, based on Gemma 3. It features a specialized tokenizer for the Ukrainian language, making it 1.5 times more token-efficient for Ukrainian text compared to the original Gemma 3. This model excels in Ukrainian language processing, including translation, summarization, Q&A, and image processing, and is designed for open research and commercial use.
Loading preview...
Lapa LLM v0.1.3-instruct: Ukrainian Language Focus
Lapa LLM v0.1.3-instruct is a 12 billion parameter instruction-tuned model built upon Gemma 3, developed by a consortium of Ukrainian AI researchers. This version includes alignment fixes and pipeline tests, maintaining the core model from v0.1.2. Its primary differentiator is its strong focus on Ukrainian language processing, aiming to be the leading open model for this language.
Key Capabilities & Achievements
- Optimized Ukrainian Tokenizer: Features a state-of-the-art tokenizer adaptation that replaces 80,000 tokens with Ukrainian ones, making it 1.5 times more token-efficient for Ukrainian text than the original Gemma 3. This results in three times fewer computations for better results.
- Efficient Instruction-Tuning: Demonstrates highly competitive performance in instruction-tuned benchmarks, closely trailing current leaders like MamayLM in some categories.
- Multimodal Support: One of the best models in its size class for image processing in Ukrainian, as measured on the MMZNO benchmark.
- Translation Excellence: Achieves 33 BLEU on FLORES for English-to-Ukrainian and Ukrainian-to-English translation, facilitating cost-effective NLP dataset translation.
- RAG System Performance: Excels in summarization and Q&A tasks, making it highly effective for Retrieval Augmented Generation (RAG) systems.
- Openness and Transparency: The project emphasizes maximum openness, providing the model for commercial use, publishing 25 datasets, disclosing data filtering methods (including for disinformation detection), and offering open-source code and training documentation.
Good For
- Processing sensitive documents in Ukrainian without external data transfer.
- Building RAG systems and chatbots that generate culturally and historically appropriate Ukrainian text.
- Developing specialized solutions through fine-tuning for specific tasks.
- High-quality machine translation between English and Ukrainian.
- Research and development in Ukrainian NLP due to its strong pretraining benchmarks and open resources.
Lapa LLM also supports image inputs (896x896 resolution, 256 tokens) and has a total input context of 128K tokens, with an output context of 8192 tokens.