Large Language Models: The Hard Parts

Foreword

In 2024, we started working on our first GenAI curriculum at Code.org. We had an ambitious goal: instead of simply introducing K-12 students to ChatGPT (or an interface that wrapped ChatGPT), we wanted to develop a small language model that they could experiment with more directly.

We believed this approach would have several benefits. First, we felt that a smaller, open source model would give us more control over hosting, pricing, and model versions. We also believed that a smaller model would hallucinate more frequently. While this might sound like an odd requirement, frequent hallucinations offer unique teaching opportunities in the classroom. Finally, we imagined a world where students could learn how to fine-tune these models using their own training data.

This approach of using small language models, however, wasn’t without its challenges. With frequent hallucinations, we had to set a high safety bar for the students, creating similar safeguards to the ones available with frontier models. We also wanted to accurately evaluate our small models to prevent regressions on the learning objectives we had set out in the curriculum. Finally, we wanted to create a framework where we could implement many of the capabilities found in more sophisticated systems, such as RAG (retrieval augmented generation) and multimodal input, but make it accessible for middle school and high school students.

It was in the middle of this work that I was introduced to Thársis Souza. Over ...