Training Azerbaijani language models on Amazon SageMaker AI
Azercell built an Azerbaijani LLM on SageMaker with 23% higher throughput and 58% lower GPU memory
Azercell Telecom partnered with AWS to train a production-ready Azerbaijani language model using Llama 3.2 1B as a foundation, achieving significant efficiency gains through Liger Kernel optimizations and a custom monolingual tokenizer. The three-stage framework (tokenizer development, continued pre-training, LoRA fine-tuning) is designed to scale to larger models. This is a useful engineering case study for teams working with morphologically complex or low-resource languages, but represents incremental applied ML rather than a major industry shift.