haolpku/K12-KGraph

What it solves

K12-KGraph addresses the lack of "curriculum cognition" in modern Large Language Models (LLMs). While LLMs can define concepts, they often struggle to understand the structured relationships between educational concepts, such as prerequisites, related experiments, and their specific location within a textbook curriculum.

How it works

The project constructs a curriculum-aligned knowledge graph (KG) from official K-12 textbooks (specifically the People's Education Press Chinese textbooks) across Math, Physics, Chemistry, and Biology.

The pipeline involves:

1. Parsing: Using MinerU to extract text from PDF textbooks.
2. Extraction: Using GPT-5.2 to extract concepts, skills, and relations based on a constrained schema.
3. Refinement: Hierarchical merging and DAG (Directed Acyclic Graph) validation combined with expert review.
4. Application: The resulting KG is used to generate two primary assets: K12-Bench (a benchmark of 23,640 multi-select questions to evaluate structural reasoning) and K12-Train (a dataset of 2,267 instruction-response pairs for supervised fine-tuning).

Who it’s for

• AI Researchers: Those working on educational LLMs or knowledge-graph-enhanced training.
• Model Developers: Developers looking to improve the structural reasoning and curriculum understanding of LLMs.
• Educators: Those interested in how structured knowledge can be used to benchmark AI's pedagogical accuracy.

Highlights

• Multi-dimensional Alignment: Aligns concepts across definitions, formulas, experiments, exercises, and structural locations.
• Expert-Verified: High quality assurance with a Fleiss' κ of 0.84 and a 98.4% correctness rate in expert spot-checks.
• Ready-to-use Assets: Provides a comprehensive benchmark (K12-Bench) and SFT training data (K12-Train).
• Pre-trained Checkpoints: Includes three standalone checkpoints (Llama 3.1, Qwen 3.5, and Qwen 3) fine-tuned on the curriculum data.