haolpku/K12-KGraph

What it solves

K12-KGraph addresses the lack of "curriculum cognition" in modern LLMs. While models can define concepts, they often struggle to understand the structured relationships between educational concepts—such as prerequisites, taxonomic neighbors, and the specific experiments or exercises that verify or test a concept.

How it works

The project builds a curriculum-aligned knowledge graph (KG) from official K-12 STEM textbooks (Math, Physics, Chemistry, Biology). The pipeline involves:

1. Extraction: Using MinerU for PDF parsing and GPT-5.2 for schema-constrained extraction of concepts, skills, and experiments.
2. Structuring: Hierarchical merging of data and DAG (Directed Acyclic Graph) validation to ensure logical flow.
3. Asset Generation: The KG is used to derive two primary assets:
   • K12-Bench: A benchmark of 23,640 multi-select questions to evaluate structural reasoning.
   • K12-Train: A dataset of 2,267 instruction-response pairs for Supervised Fine-Tuning (SFT) to teach models curriculum structure.

Who it’s for

AI researchers and developers working on educational LLMs, curriculum-aware AI, and those interested in benchmarking structural reasoning in STEM subjects.

Highlights

• Multi-Dimensional Alignment: Aligns concepts across definitions, formulas, experiments, exercises, and textbook locations.
• High Quality: Validated by subject experts (Fleiss' κ = 0.84) and automatic DAG validation.
• Ready-to-use Models: Provides three fine-tuned checkpoints (Llama 3.1, Qwen 3.5, Qwen 3) specifically trained on curriculum data.
• Comprehensive Schema: Includes 7 node types (e.g., Concept, Skill, Experiment) and 9 edge types (e.g., prerequisites_for, verifies) to map educational dependencies.