← Back to blog
AiAbout 6 min read

A Chinese K-12 Knowledge Graph With 1,644 Nodes and the Edges That Make It Useful

Published Nov 13, 2026
A Chinese K-12 Knowledge Graph With 1,644 Nodes and the Edges That Make It Useful

A repository called cn-math-taxonomy picked up roughly eight thousand stars since it appeared on October 18, and the name undersells it badly. It started as a map of Chinese school mathematics and now covers English from grade three upward, middle school physics, and middle school biology.

The numbers are the reason people star it. 1,644 knowledge points. 2,403 prerequisite edges. 6,723 questions. Aligned, chapter by chapter, to the 2022 curriculum standards as revised in 2025.

What is actually inside

The graph is built from the textbooks rather than from a syllabus summary. Mathematics follows the People's Education Press series for grades one through nine, plus the 2019 A edition with two compulsory volumes and three elective ones for high school. English follows PEP from grade three, the Go for it! series for middle school, and the 2019 high school set. Physics separately covers the eighth and ninth grade volumes and the high school compulsory and elective sequences.

That choice, textbook rather than standard, is what makes the thing usable. A curriculum standard says a student should understand linear functions by a certain grade. A textbook says chapter nineteen introduces them after chapter eighteen on quadratic equations, and that ordering is what a diagnostic engine needs.

The structure has some deliberate decisions in it that are worth reading closely. English is a separate graph and does not connect to mathematics at all. Middle school English connects back to primary English, and high school English connects mainly back to middle school English. Physics is its own graph too, and the mathematics a physics concept depends on is stored in a mathLinks field rather than as a prerequisite edge. The author's reasoning is implicit but sound: a physics student who cannot solve a quadratic is not missing a prerequisite in the graph-theoretic sense, they are missing a tool, and conflating the two would make the diagnostic walk into the wrong subject.

The diagnostic is the product

The headline feature is adaptive prerequisite diagnosis, and it works the way the graph implies.

A student gets a high school question wrong. The system walks backwards along the prerequisite edges, checking the concepts underneath, and keeps walking until it reaches a specific primary school knowledge point that is actually the root of the failure. Not "you need to review functions." Something closer to "you are shaky on the relationship between fractions and division, which is why the rational expression simplification failed."

That is a different product from a question bank with difficulty tags. It is closer to how a good tutor thinks, and it only works if the edge set is complete enough to walk on. With 1,644 nodes and 2,403 edges, the graph averages under 1.5 outgoing edges per node. That is sparse. Sparse is correct for a prerequisite DAG, since most concepts depend on one or two things, and it also means a single wrong edge sends the walk somewhere useless.

The part I did not expect

The README contains something rare in open data projects: an explicit list of what could not be verified.

The author went looking for the official table of contents for the 2024 edition of the ninth grade second-semester mathematics textbook and could not find one. The publisher's introduction page describes the editorial approach without listing chapters, and the page titled for that volume shows an old cover. The 2024 edition is not expected to enter classrooms until spring 2027.

So the repository labels the chapter list as "the most reliable inference, not the official text," cites the third-party teaching-aid page it drew from, and tells readers to check the physical book. It notes where a chapter name changed, from a section on probability to one on random events, and flags which textbook editions are not yet published and are being substituted with the previous version.

Most projects at this scale fill those gaps and move on. Documenting them costs nothing and saves every downstream user from building on a guess they did not know was a guess.

Licensing, before you build a business

The code and data are under CC BY-NC 4.0. Personal study, teaching, and research are free. Paid courses, paid apps, and any service where an institution charges money require separate authorization, which the author handles directly by message or email.

For anyone outside China planning to use this, that clause is the whole conversation. A non-commercial license on a knowledge graph is fine for a research prototype and a blocker for a product. The data itself is also jurisdiction-specific in a way that is hard to port: the edge structure encodes the order in which Chinese textbooks introduce concepts, and that order does not match how a US or UK curriculum sequences the same material.

Why this matters beyond China

Two reasons, and neither is about education technology as a market.

First, it is a clean example of the thing knowledge graphs are actually good for. Constraint following. A prerequisite DAG lets you answer "what does this student need to know before this question makes sense," and that question has a definite answer that a language model will happily approximate and occasionally get wrong. The graph does it in a traversal.

Second, it shows where the value sits in a data product. The 6,723 questions are commodity. The 1,644 nodes are a week of careful work. The 2,403 edges, and the decisions about which dependencies are edges and which are fields, are the part that took real judgment. Anyone can generate questions now. Almost nobody sits down and argues with themselves about whether a physics concept's mathematical dependency belongs in the graph.

It also reframes what a question bank is worth. Six thousand questions sound like the hard part, and ten years ago they were. Now a model can produce questions by the thousand, tagged by topic, in an afternoon. What it cannot produce is the ordering: the claim that this concept depends on that one, made by someone who has taught both and knows where students actually break. That claim is the scarce input, and it is the thing a non-commercial license is protecting.

If you are evaluating the graph for your own use, one test tells you most of what you need. Pick a concept near the top of a subject, walk its prerequisites down to the bottom, and read the chain out loud. If each step sounds like something a teacher would say, the edges were built by someone paying attention. If a step jumps from factoring quadratics to understanding negative numbers with nothing in between, you have found where the graph runs out of evidence and starts inferring.

If you want to try it, the repository and the demo site at math.pansgo.com are the entry points, and the data is plain enough to load into anything that speaks graphs. Read the section on what is provisional first. That is the part that tells you how much to trust the rest.

Related articles