DSML.KZ Community Hub

Til-Qazyna builds digital and AI infrastructure for the Kazakh language, including foundational language resources and open-source models. The team develops linguistic tools, custom text corpora, morphological analyzers, and fine-tuned large language models. Responsibilities: • Fine-tune, evaluate, and benchmark open-source LLMs for Kazakh using LoRA, QLoRA, and PEFT. • Develop and optimize rule-based and neural NLP components, including morphological analyzers, lemmatizers, tokenizers, and terminology parsers. • Build pipelines for text crawling, cleaning, deduplication, synthetic data generation, and dataset curation. • Package NLP models into Docker and FastAPI microservices and optimize inference with vLLM, Ollama, TensorRT-LLM, or ONNX. Requirements: • 1+ years of hands-on NLP, AI, or machine learning engineering experience. • Understanding of Transformers, attention mechanisms, embeddings, and tokenization; experience with PyTorch, Hugging Face libraries, and Scikit-learn. • Experience building preprocessing pipelines for structured and unstructured text data. • Practical knowledge of Docker, Linux, and Git. • Degree in Computer Science, Data Science, or Applied Mathematics. Optional: • Experience with LLM serving and optimization, including vLLM, TGI, llama.cpp, AWQ, or GPTQ. • Experience with RAG stacks such as LangChain, LlamaIndex, ChromaDB, Qdrant, or FAISS. • Understanding of Kazakh morphological, agglutinative, or grammatical characteristics; familiarity with DPO, RLHF, or synthetic data generation.