VideoEduBench Dataset

This dataset contains expert-annotated multimodal interaction events for VideoEduBench, a diagnostic benchmark for evaluating multimodal large language models (MLLMs) on long-horizon classroom interaction understanding.

📄 Paper: VideoEduBench: Diagnostic Benchmarking of Video Agents for Long-Horizon Classroom Interaction Understanding (KDD 2026 Undergraduate Consortium)

🔗 Code repository: https://github.com/tinaxie123/VideoDR-Benchmark

📊 Dataset Statistics

• 62 classroom lessons (~40 hours of HD video, 1080p / 30fps)
• 10,552 expert-annotated multimodal interaction events
• Cohen's κ = 0.83 inter-rater agreement on 10% double-annotated subset
• Subject coverage: 30 Science / 26 Liberal Arts / 6 other lessons
• Grade levels: 55 elementary, 7 secondary

🗂️ Files

File	Description
1.xlsx ~ 62.xlsx	Per-lesson annotation files (one row per multimodal interaction event)
all_62_cases_long.xlsx	Aggregated long-format dataset across all 62 lessons

📋 Annotation Schema (MCDCS)

Each event is annotated using the Multi-agent Classroom Dialogue Coding System (MCDCS):

Field	Type	Description
timestamp	float	Time in seconds from lesson start
agent	string	One of: Teacher / Student / AI
behavior	string	Fine-grained behavior code (~24 categories)
intent	string	Pedagogical intent: Instructional Guidance / Evaluation / Dynamic Presentation / Content Generation
modality	string	visual / audio / multimodal

🎯 Dual-Path Evaluation Paradigm

The benchmark introduces a dual-path diagnostic evaluation:

1. Fine-grained Multimodal Perception (FMP): Tests whether models can localize and identify specific behavioral codes (gestures, screen operations, audio feedback) over the full 40-minute context.
2. Pedagogical Logic Reasoning (PLR): Probes deep instructional inference — e.g., distinguishing a teacher's deliberate wait-time strategy from a technical glitch.

📐 Evaluation Metrics

• ISA (Interaction Sequence Accuracy): Perception fidelity
• PI-F1 (Pedagogical Intent F1): Reasoning quality
• ARS (Agentic Retrieval Success): Long-horizon tool-use, with 2-second temporal tolerance

🎥 Video Access

The 62 raw classroom video recordings are hosted externally due to size and original platform licensing. Access instructions are provided in the GitHub repository.

🚀 Quick Start

import pandas as pd

# Load aggregated dataset
df = pd.read_excel("all_62_cases_long.xlsx")

# Or load per-lesson files
lesson_1 = pd.read_excel("1.xlsx")
print(lesson_1.head())

📄 Citation

@inproceedings{xie2026videoedubench,
  title     = {VideoEduBench: Diagnostic Benchmarking of Video Agents for Long-Horizon Classroom Interaction Understanding},
  author    = {Xie, Haotong},
  booktitle = {KDD 2026 Undergraduate Consortium},
  year      = {2026}
}

📜 License

• Annotations (this repository): CC BY-NC 4.0
• Original videos: Governed by the Socrates platform's licensing terms
• Code: MIT License (in the GitHub repository)

📧 Contact

Haotong Xie · Shanghai University of Finance and Economics
📧 haotongxieqaq@163.com
🔗 GitHub Profile · HuggingFace Profile