Hugging Face's logo

Datasets:
aigrant
/
awesome-taiwan-knowledge

Tasks:
Table Question Answering
License:
Dataset card Viewer Files Community
2
awesome-taiwan-knowledge / README.md
gatelynch's picture
gatelynch
update the evaluation part
effcebe verified
preview code
|
raw
history blame contribute delete
5.74 kB
metadata
license: cc-by-4.0
task_categories:
  - table-question-answering
configs:
  - config_name: default
    data_files:
      - split: civics_studies_hs
        path: ATK_August_2024/mcq_civics_studies_hs.csv
      - split: social_studies_elem_jhs
        path: ATK_August_2024/mcq_taiwan_social_studies_elem_jhs.csv
      - split: mtqs_sicial_studies_elem_jhs
        path: ATK_August_2024/mtqs_sicial_studies_elem_jhs.csv
      - split: mtqs_taiwan_literature
        path: ATK_August_2024/mtqs_taiwan_literature.csv

Awesome Taiwan Knowledge (ATK) Dataset

The Awesome Taiwan Knowledge (ATK) Dataset is a comprehensive collection of questions and answers designed to evaluate artificial intelligence models' understanding of Taiwan-specific information. This unique dataset addresses the growing need for culturally nuanced AI performance metrics, particularly for models claiming global competence.

Key Features:

  1. Taiwan-Centric Content: Covers a wide range of topics uniquely relevant to Taiwan, including history, culture, politics, education, and current affairs.

  2. Diverse Question Formats:

    • Multiple-choice questions for quantitative assessment
    • Multi-turn dialogue questions to evaluate contextual understanding and conversational abilities

  3. Expert-Validated Answers: All responses are meticulously curated and verified by qualified Taiwanese educators and subject matter experts.

  4. Detailed Explanations: Each question is accompanied by in-depth explanations, providing context and educational value beyond mere right/wrong evaluations.

  5. Continuous Updates: The dataset is regularly refreshed to include current events and evolving cultural nuances.

Focused Subject Areas:

The ATK Dataset collects questions from key educational domains, ensuring comprehensive coverage of Taiwan-specific knowledge:

  1. Civic Studies for High School
  2. Social Studies for Elementary School and Junior High
  3. Taiwan Literature for K-12
  4. Taiwan Geography
  5. Taiwan History

These areas represent core components of Taiwan's educational curriculum, providing a robust foundation for assessing AI models' understanding of Taiwan's societal, cultural, and geographical landscape.

Purpose:

  • Benchmark AI models' proficiency in Taiwan-specific knowledge
  • Identify gaps in AI systems' understanding of localized information
  • Promote the development of more culturally aware and inclusive AI models
  • Provide a standardized tool for comparing different AI models' performance on Taiwan-related queries

Current Status:

The ATK Dataset is in active development, with ongoing data collection from local educators and experts. A comprehensive benchmarking report, evaluating various AI models against this dataset, is forthcoming.

Significance:

This dataset aims to highlight the importance of cultural and regional knowledge in AI systems, encouraging developers to create more inclusive and globally competent models. By focusing on Taiwan-specific information, the ATK Dataset addresses a critical gap in current AI evaluation metrics.

Evaluation:

Here's the table using claude as the evaluation model to see how GPT-4o, Claude sonnet 3.5 and Gemini perform on answering the questions:

Model Subject (1) Overall Model Response Accuracy (2) Model Response Confidence Average (0-100) (3) Model Response Key Confidence Average
GPT-4o Overall 70.35% 75.11 63.61
Elementary School Civics Studies 94.00% 76.52 65.00
High School Taiwan Literature 78.89% 83.52 74.21
High School Society Studies 26.32% 71.20 66.71
Junior High Society Studies 69.29% 67.16 58.57
Claude 3.5 Sonnet Overall 53.76% 85.44 63.13
Elementary School Civics Studies 67.33% 85.20 50.20
High School Taiwan Literature 50.00% 87.22 67.56
High School Society Studies 21.05% 67.60 81.20
Junior High Society Studies 49.61% 84.84 48.75
Gemini Output Overall 32.68% 83.34 32.54
Elementary School Civics Studies 47.33% 79.65 21.19
High School Taiwan Literature 44.44% 85.13 36.10
High School Society Studies 8.42% 91.25 43.91
Junior High Society Studies 25.98% 84.70 29.79

Key Observations:

  • Subject-wise Performance:

    • Elementary School Civics Studies: All models performed relatively well here, with GPT-4 leading (94%), followed by Claude (67.33%), and Gemini (47.33%)
    • High School Taiwan Literature: GPT-4 showed strong performance (78.89%), while Claude (50%) and Gemini (44.44%) were notably lower
    • High School Society Studies: All models struggled here, with particularly low accuracy (GPT-4: 26.32%, Claude: 21.05%, Gemini: 8.42%)

  • Confidence Levels:

Interestingly, all models showed relatively high confidence (mostly above 70%) despite varying accuracy levels Claude 3.5 Sonnet and Gemini often showed higher confidence than GPT-4, despite lower accuracy This suggests potential overconfidence issues, particularly in Gemini and Claude

  • Strongest Areas:
    • GPT-4: Elementary School Civics Studies (94%)
    • Claude: Elementary School Civics Studies (67.33%)
    • Gemini: Elementary School Civics Studies (47.33%)

Contributors

年級 領域 教師名稱 學校
小學 公民 朱堯麟 退休
國中 台灣文學 陳雅娟 竹北國中
高中 公民 廖宗德 六家高中
and 5 more annonymous contributors