AcupunctureEval: A Multimodal Benchmark for Evaluating Generative Models in the Acupuncture Domain

🌈 Updates • 📖 Overview • 🥸 Dataset

📊 Evaluation • 📄 License

中文 | English

The complete dataset can be accessed on: https://huggingface.co/datasets/hf-anon-9-1/AcupunctureEval

🌈 Updates

• AcupunctureEval website is online: https://AcupunctureEval.daytime001.xin
• AcupunctureEval platform initialized, supporting multi-dimensional VQA and QA evaluation.

📖 Overview

AcupunctureEval is the first large-scale multimodal benchmark specifically designed for evaluating generative models in the acupuncture domain. It covers text, image, and video modalities, spanning multidimensional knowledge including acupuncture theory, meridians and acupoints, clinical reasoning, TCM syndromes, and practical acupoint operation.

• First Acupuncture-Specific Benchmark: Entirely self-constructed without relying on existing public datasets
• Multimodal Coverage: Integrates text (AcupunctureQA), visual QA (AcupunctureVQA), and video (AcupunctureVideo)
• Large-Scale: Contains 58,686 question samples and 1,216 video samples
• Diverse Task Types: Objective questions, subjective QA, case analysis, image understanding, image reasoning, and video understanding
• Human-Machine Collaborative Pipeline: Ensures both data quality and scale through expert validation

📊 Evaluation Channel

We provide online evaluation submission methods:

1. Web Interface Submission: Upload the prediction result JSON file on the "Evaluation" page.

Prediction File Format Description:

• Please refer to the JSON structure in 'Reference Case'.
• Flat question types (A1/A2/X) submit array format.
• Grouped question types (A3/A4/B) need to include parent question ID and sub-question outputs.

🥸 Dataset Introduction

Components

The dataset of this project covers multiple knowledge dimensions in the field of Acupuncture and Traditional Chinese Medicine (TCM), aiming to comprehensively evaluate the professional capabilities of large models.

1. QA (Text Question Answering) - AcupunctureQA

Comprehensive and multi-level evaluation of the model's TCM theoretical knowledge, containing three parts: objective questions, subjective questions, and case analysis:

• Objective Questions: A total of 1735 samples, sourced from textbooks.
  • Covered Fields: Meridians and Acupoints, Acupuncture and Moxibustion, TCM Diseases.
  • Dataset Split: Validation Set (1516) + Test Set (219).
  • Question Types: A1, A2, A3, A4, B, X types.
• Subjective Questions: A total of 45,962 question-answer pairs.
  • Covered Fields: Acupuncture (5205), Acupoints (12576), Moxibustion (644), Tuina (1416), TCM Syndromes, etc.
• Case Analysis: A total of 100 samples, covering 7 dimensions such as etiology and pathogenesis.

2. VQA (Visual Question Answering) - AcupunctureVQA

Evaluates the model's multimodal understanding capabilities regarding acupuncture points, anatomical structures, and operational techniques, with a total of 10,729 samples:

• Task Types: Image Understanding and Image Reasoning.
• Question Classification:
  • Single/Multiple Choice: TCM diagnosis and identification based on images.
  • Localization: Precise identification of acupoints and body parts.
  • Operation: Visual question answering on acupuncture techniques and operation procedures.

3. Video (Video Understanding) - AcupunctureVideo

A video dataset regarding practical acupuncture point operations, with a total of 1000 samples, divided into three categories:

1. Standard Techniques: Standard practical operation of acupuncture points on the human body by acupuncturists.
2. Clinical Records: Actual operation videos on patients by experts in real clinical scenarios.
3. Animal Experiments: Standard practical operation of acupuncture points on experimental pigs by acupuncturists.

Data Examples

QA Item Example (Type A1)

{
    "ID": "1",
    "question": "The earliest and relatively complete monograph on acupuncture and moxibustion is ( )",
    "options": [
        "A. The AB Classic of Acupuncture and Moxibustion",
        "B. Miraculous Pivot",
        "C. Essentials of Acupuncture and Moxibustion in the Bronze Man",
        "D. Classic of Nourishing Life with Acupuncture and Moxibustion",
        "E. Elucidation of the Fourteen Meridians"
    ],
}

VQA Item Example (Image Understanding)

{
    "ID": "1",
    "Type": "Image Understanding",
    "Class": "Acupoints of Lung Meridian of Hand-Taiyin",
    "Images": [
        "Figure 3-2-2.jpeg"
    ],
    "Question": "Which acupoints in the options might be included in the image?",
    "Options": [
        "A. Lieque",
        "B. Yunmen",
        "C. Kongzui",
        "D. Jingqu"
    ],
}

📄 License

This project follows the MIT License.