Smol Course is a practical educational resource for learning to align language models for specific use cases.
The course addresses the challenge of customizing language models without requiring extensive computational resources or paid services. It teaches alignment techniques through hands-on exercises built around SmolLM3 and SmolVLM2 models, with minimal GPU requirements. The skills covered—instruction tuning, evaluation, preference alignment using algorithms like DPO, and vision language model adaptation—apply to both small and larger models. The course emphasizes that small language models suit domain-specific applications where efficiency, customization, local deployment, and cost control matter more than general-purpose capabilities.
Developers should choose this course if they want to fine-tune models on local machines without cloud dependencies or significant hardware investment. It works well for teams building domain-specific applications, academic researchers exploring cutting-edge techniques with lower logistical overhead, and anyone prioritizing privacy by keeping data local. The course assumes basic familiarity with machine learning, Python, PyTorch, and the transformers library. It is structured as a peer-reviewed, community-driven resource where learners submit work for review and can contribute models to a leaderboard.
The project maintains active engagement through an open pull request model where contributors read material, complete exercises, add examples, and submit models for community evaluation. The course roadmap shows completed modules on instruction tuning, evaluation, and preference alignment, with additional modules on reinforcement learning and synthetic data generation planned for future release.
