Training
1 min read
One course, two audiences: training people and models together
The same clear lesson that helps a person learn can also become clean training data for a model. Here is why we publish every course both ways.
Every course on this site is written once and served two ways: as readable lessons for people, and as structured data for machines. That is not just convenient. Good teaching material and good training data turn out to have a lot in common.
What makes material good for both
- Clear structure. Headings, short sections and one idea at a time help a reader follow along, and help a model learn clean boundaries between concepts.
- Accurate, reviewed content. Mistakes confuse learners and get baked into models. Quality beats quantity in both cases.
- Questions with explanations. Quizzes test people. The same question, answer and explanation make a natural instruction-style example for fine-tuning.
Two dataset formats
Through the API, each course can be downloaded as JSONL:
- documents: one record per lesson, suited to retrieval (RAG) or further pretraining.
- instruct: chat-style message pairs built from lessons, key points and quiz questions, suited to fine-tuning.
Licensing matters
Every course carries a license, CC-BY-4.0 by default, so you know how its content may be reused. Check it before you train on a dataset.
To get started, visit For machines to create an API key and browse the available datasets.