Happy-LLM is a free, open-source educational project from Datawhale that teaches the principles and training process behind large language models. It begins with NLP foundations and progressively explains attention, Transformer architecture, pretrained language-model families, modern LLM design, training strategies, and the capabilities that emerge as models scale.

The course combines theory with implementation. Learners build a small LLaMA2-style model, train tokenizers, experiment with pretraining and supervised fine-tuning, use parameter-efficient techniques such as LoRA and QLoRA, and study practical LLM applications including evaluation, RAG, agents, and newer agentic reinforcement-learning approaches such as GRPO and Search-R1-style workflows.

Happy-LLM is aimed at students, researchers, and AI enthusiasts who already have some Python experience and ideally basic deep-learning knowledge. Unlike tutorials focused only on calling model APIs, it tries to connect mathematical and architectural ideas to executable code, giving learners a path from understanding how LLMs work to training small models and building applications around them.