A Chinese Curriculum Became 870 Math Nodes. The Best Part Is the List of What Not to Trust.

A repository called cn-math-taxonomy has 7,058 stars, and it does not contain a model. It contains a graph.
The graph is the Chinese school curriculum, decomposed. For math: 284 nodes for primary school grades 1 to 6, 315 for junior high, 143 for the compulsory high school volumes, 128 for the elective ones. That is 870 knowledge points and 1,459 prerequisite edges inside math, plus 145 more that cross stages, junior high to primary, compulsory to elective, module to module.

There are also 314 curriculum-standard entries and 3,627 questions.
The English graph is separate by design: 275 + 257 + 259 nodes with 1,028 edges, deliberately not connected to math. Physics adds 92 and 177. Chinese language adds 493 and 189. The total question bank is around 10,600.
What the graph is for
The headline feature is an adaptive prerequisite diagnosis. A student gets a math question wrong, and the system walks the prerequisite edges backward, from the high school concept, through junior high, down to the specific primary school node that is probably the actual gap.
The demo in the README makes it concrete. A student fails a question on completing the square, and the diagnosis traces back to the perfect square formula. A high school question on the concept of a function traces back to quadratic functions in junior high.
That is the thing a language model is bad at, and the reason a graph is worth building. A tutor that only knows the current question can re-explain the current question. A tutor that knows the prerequisite chain can tell you the misunderstanding started three years earlier, in a specific, named unit.
The data model is borrowed: `topics` plus `dependencies`, following the shape of Marble's os-taxonomy.
Provenance is the actual engineering
The README's opening paragraphs are a citation list. The math graph is aligned to the 2022 compulsory education math standards as revised in 2025, and to the 2017 high school standards as revised in 2020 and 2025, covering the compulsory and elective-compulsory volumes. The textbook edition is Renjiao A for high school, the 2019 printing, three compulsory volumes and three elective.
Then the document does something unusual. It states, in the same breath as the counts, which parts it could not verify.
Grade 9 second semester uses a table of contents assembled from third-party sources after the official one could not be found. Renjiao's product page still lists the 2014 first edition. The 2024 revision of that volume is expected to enter use in spring 2027, so the chapter list is a "most reliable guess", labelled as unofficial and pending verification.
The curriculum standard quotes are paraphrases, not verbatim text. The repository records printing page numbers from a reposted PDF so that a reader can check the source, and it marks the sentence text as `verbatim: false`.
The junior high textbook chapter lists were checked against the national smart education platform page previews, with the book numbers recorded. The primary school volume tables were only spot-checked.
The prerequisite edges themselves carry the strongest warning. Which relationships are hard, which are soft, and whether a specific edge should exist at all, the README says, need review by practising teachers before any of this is used for teaching, diagnosis or learning-path recommendation.
The question that matters most
Here is the line that makes the project trustworthy: the high school and English questions have not been blind-solved. The verification script and the answer come from the same expression, so a check that passes proves the two agree, not that either is right.
Primary and junior high math did get independent verification. Junior high math also got a blind-solve pass. Primary math has a separate program re-checking it.
So the repository grades its own confidence by section, and the grade correlates with exactly the thing a tutoring product would depend on. If you are going to tell a student they have a gap, the question that revealed the gap had better be correct.
Almost no dataset ships that distinction. Most ship a row count and a licence.
How you would use it
The data is JSON, the repo is Python, and there is a hosted demo for the diagnosis flow. The commercial terms are the constraint: CC BY-NC 4.0, free for personal study, teaching and research, with paid courses, paid apps and institutional paid services requiring a separate licence.
That non-commercial clause is the honest consequence of the sourcing. The curriculum standards are government documents, but the alignment work, the question banks and the prerequisite judgements are the author's, and the author has priced commercial use separately.
There is also a practical warning embedded in the file structure. Fractional grade labels like 10, 11, 12, 13 for high school volumes are book labels, not teaching schedules. The README says plainly that this is not a statement about pacing and that teachers should adjust it.
Why the graphs are kept apart
One structural decision in the repository is easy to skip past: math, English, physics and Chinese each get their own graph, and the edges do not cross.
Physics references math through a separate `mathLinks` field rather than a prerequisite edge. English is its own graph entirely, connecting within itself from high school back to primary. The README states the reasoning in a line: physics uses math, but that is a relationship, not a prerequisite.
The distinction is the difference between a graph that can be trusted and one that looks impressive. Saying a physics lesson requires calculus is defensible. Saying it requires a specific math node in a specific grade is a claim about teaching order that varies by school, by teacher and by province. Drawing that edge would inflate the edge count and make the diagnosis worse.
The same restraint shows up in the question banks. English deliberately has no listening or oral items, because a text dataset cannot test them honestly. The result is a graph with fewer edges than a maximal one and a smaller claim that it can actually support.
Why this matters beyond one repo
The AI tutoring market has spent two years discovering that a good language model is not a curriculum. Ask a model to explain quadratic functions and it will do a competent job. Ask it which prerequisite a specific student missed, and it will produce a plausible answer with no provenance.
A prerequisite graph is the missing layer, and building one is unglamorous work: reading standards, cross-checking textbook editions, arguing about whether one concept really depends on another. The 7,000 stars suggest people recognise how rare that work is, especially published with its own defect list.
The second thing the repository demonstrates is a template for how to release data like this. Record the source edition, record the printing page, distinguish a quote from a paraphrase, name the sections that were blind-checked and the sections that were not, and say which decisions still require a professional. A dataset with that documentation can be adopted by cautious people. A dataset without it can only be adopted by people who will not check.
The thin part
The graph is Chinese-curriculum specific, so its direct reuse outside China is limited. The method transfers; the 870 nodes do not.
It is also a snapshot of a curriculum in motion. The standards were revised in 2025, some textbook volumes have not shipped, and one chapter list is explicitly a guess. The repository dates its checks, which is the right move and also a reminder that this is a document that decays.
For anyone building tutoring tools, evaluation sets or curriculum-aware agents, the value is the pattern rather than the file: encode the dependency structure, record provenance down to the page, and publish the parts you could not verify, because those are the parts the next person needs before trusting the source.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.