Hi everyone,<\/p>\n
Sharing a new open dataset for people working on education AI, curriculum modeling, or LLM training.<\/p>\n
K12-KGraph is a curriculum-aligned knowledge graph built from publicly available K-12 textbook materials. The current release covers math, physics, chemistry, and biology, and includes structured links between concepts, skills, experiments, exercises, textbook sections, chapters, and books.<\/p>\n
The main idea is simple: for education LLMs, adding more practice questions is useful, but it often only teaches the model how to answer questions. A curriculum knowledge graph can also teach the model how topics are connected, which concepts should come first, and what knowledge may be missing when a student gets stuck.<\/p>\n
The released resources include:<\/p>\n
- \n
- A curriculum knowledge graph<\/li>\n
- A benchmark for testing curriculum understanding<\/li>\n
- A prepared training dataset generated from the graph<\/li>\n
- The paper and construction method, so the same approach can be adapted to other textbook systems where content rights are clear<\/li>\n<\/ul>\n
In the experiments, the graph-based training data performed better than the same amount of regular instruction or exercise-style data on education benchmarks. The useful takeaway is that structure matters: modeling the curriculum itself can improve education LLMs more efficiently than only scaling question banks.<\/p>\n
Links:<\/p>\n
Paper + Dataset: https:\/\/huggingface.co\/papers\/2605.09635<\/a><\/p>\n<\/div>\n
submitted by \/u\/Puzzleheaded_Box2842 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
