LLM Training Data (Education)

Product Overview

This product is an Education & Pedagogy LLM Training Data Package covering K-12 + Higher Education multi-subject content, including high-quality tutoring dialogues, instructional reasoning chains, exam question banks, curriculum knowledge graphs, and pedagogical strategy annotations. Directly applicable for LLM fine-tuning (SFT), RLHF preference alignment, and building education-focused RAG systems.

• Total Data: ~15M+ high-quality educational data entries (dialogues, QA, reasoning chains)
• Token Scale: ~9.5 billion tokens
• Grade Coverage: Elementary → Middle School → High School → Undergraduate → Graduate/Professional Exams
• Subject Coverage: Mathematics, Physics, Chemistry, Biology, Language Arts, History, Geography, Computer Science, and 15+ subjects
• Data Format: JSONL / Parquet (with structured field annotations)
• Quality Standards: Pedagogical intent annotations, difficulty grading, knowledge point mapping, deduplication & benchmark contamination detection

---

Data Types & Content

---

Data Quality & Features

• Pedagogical Intent Annotation: Following ACL 2025 research, each tutoring turn annotated with fine-grained pedagogical intent (11 category taxonomy)
• Socratic Guidance Markup: Specifically annotated guided teaching segments for question-based discovery learning
• Knowledge Point Graph Mapping: Each entry linked to specific knowledge points (KP) for retrieval and adaptive recommendations
• Difficulty Grading System: Cognitive level annotations following Bloom's Taxonomy (Remember → Understand → Apply → Analyze → Evaluate → Create)
• Precise Grade-Level Annotation: Each entry tagged with applicable grade levels
• Benchmark Decontamination: Filtered against TutorBench, KMP-Bench education evaluation datasets
• Multi-Stage Deduplication: Exact → fuzzy → semantic three-layer pipeline

---

Compatible Education AI Evaluations

• TutorBench (ICLR 2026): 1,490 expert-annotated samples testing adaptive explanations, actionable feedback, and hint generation
• KMP-Bench: K-8 mathematical pedagogy evaluation with dialogue and skills modules
• MathDial: Math tutoring dialogue evaluation benchmark
• MMLU / ARC: Multi-subject subsets of general knowledge evaluations

---

Use Cases

• AI Tutoring Systems: Train intelligent Socratic tutoring AI (cf. Khan Academy / Khanmigo methodology)
• Adaptive Learning Platforms: Knowledge graph-based personalized learning path recommendations
• Automated Test Generation & Grading: Generate high-quality exams with detailed rubrics
• Teacher Assistance Tools: Auto-generate lesson plans, curriculum designs, differentiated instruction
• Learning Diagnostics: Student weakness diagnosis and personalized remediation plans
• Educational Research: Pedagogical effectiveness analysis, learning behavior research

---

Delivery Information

• Data Files: JSONL / Parquet (as agreed), with README and field descriptions
• Delivery Time: Typically 3-7 business days
• Data Updates: Incremental update service available (new curriculum standards/textbook adaptations)
• Compliance: No student personal privacy data; for special compliance (e.g., COPPA/FERPA), communicate before ordering