TCO-Dataset: Traditional Chinese Opera Understanding Dataset

Overview

TCO-Dataset is a comprehensive, bilingual multimodal benchmark designed to evaluate Large Multimodal Models (LMMs) on their understanding of Traditional Chinese Opera. The dataset facilitates fine-grained cultural, visual, and symbolic reasoning across eight major genres, providing a specialized resource for culturally-grounded vision-language research.

Data Source and Provenance

To ensure data consistency and long-term traceability, all data samples in this dataset are managed and released under the following institutional framework:

• Institutional Archive: All images and associated annotations are sourced from the XJTU Opera Research Database (XJTU-ORD Archive).
• Curation: This archive is an internal repository maintained by the School of Automation Science and Engineering at Xi'an Jiaotong University. It integrates original field photography captured by the authors' research teams with curated institutional digital assets.
• Data Integrity: Every sample in this dataset has undergone a rigorous quality control process to ensure high visual fidelity and the absence of third-party watermarks or branding, making it suitable for high-standard academic evaluation.

Licensing and Usage

• License: The TCO-Dataset is distributed under the Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0) license.
• Restriction: This license permits the data to be freely used, shared, and adapted for non-commercial research and educational purposes, provided appropriate credit is given to the original authors and the XJTU-ORD Archive.
• Reasoning: The CC-BY-NC 4.0 framework is applied to protect the cultural heritage and the intellectual property of the original operatic performances represented in this archive.

Dataset Structure

The dataset follows a modular structure where each sample ID (1-1000) corresponds to a set of files:

• Image (.png): High-resolution visual data.
• Question (.md): Multiple-choice question with four options.
• Ground Truth (.txt): Correct answer key.
• Metadata (.json): Consolidated metadata file.

Metadata Example (JSON)

The .json file for each sample consolidates the textual content and ground truth for easier programmatic access. Note that provenance and licensing information is declared globally in this README and applies to every sample.

{
    "id": "920",
    "image": "920.png",
    "question_file": "920.md",
    "answer_file": "920.txt",
    "question_text": "观察这位净角演员的鬓边，插着一朵显眼的黄花。在戏曲中，这种粗犷外表配戴花朵的装扮通常用来表现人物具有什么样的性格特质（如李逵）？",
    "options": {
        "A": "粗中有细、憨厚可爱或孝顺",
        "B": "阴险狡诈、口蜜腹剑",
        "C": "温文尔雅、风流倜傥",
        "D": "神机妙算、沉稳老练"
    },
    "answer": "A",
    "question_en": "Observe the yellow flower pinned to the side of the Jing actor's head. In opera, what personality trait does this contrast of a rough appearance with a flower typically suggest (often associated with characters like Li Kui)?",
    "options_en": {
        "A": "Rough but gentle, naive, cute, or filial",
        "B": "Treacherous and hypocritical",
        "C": "Gentle and romantic",
        "D": "Wise and calculating"
    }
}

Opera Genres

• Peking Opera (Jingju)
• Yue Opera (Yueju)
• Huangmei Opera (Huangmeixi)
• Cantonese Opera (Yueju)
• Sichuan Opera (Chuanju)
• Qinqiang Opera (Qinqiang)
• Yu Opera (Yuju)
• Kunqu Opera (Kunqu)

Dataset Statistics

• Total Samples: 1000
• Language: Bilingual (approx. 66% English, 33% Chinese).
• Encoding: All text-based files (.md, .json, .txt) are encoded in UTF-8.

Quality Control

• Expert Annotation: Questions and answers were authored and cross-verified by Chinese opera domain experts.
• Consistency Checks: Three rounds of internal audits were conducted to ensure the accuracy of labels and the cultural fidelity of the reasoning tasks.