How to Teach a Small LLM to Suggest K12 Creative Project Ideas
SMRTR summary
What if you could teach a small AI model to suggest hands-on cultural activities for kids, using nothing more than a laptop and a Wikipedia rabbit hole? That's exactly what one developer set out to do while building an educational app for K-12 classrooms.
Educators testing the prototype kept asking for one thing: a way to search for project ideas based on materials they already have. So the developer trained a lightweight language model on thousands of Wikipedia articles covering arts and traditions from around the world, filtering the corpus down from roughly 19,000 articles to about 6,600 relevant ones.
Using a technique called LoRA, which fine-tunes a model without retraining it from scratch, the whole process ran on a personal Apple Silicon laptop.
The result: a feature where teachers type in their available materials, and the app returns culturally grounded STEAM activities, complete with instructions and learning objectives.
Child safety guardrails were built in at every stage, screening both the training data and live outputs before anything reaches a classroom.
SMRTR provides this summary for quick context. The original article belongs to Daily.dev.
Read the original article