README.md

metadata

license: cc
language:
  - en
  - ko
tags:
  - benchmark
  - culture
  - korean
  - evaluation
  - multilingual
  - superstition
pretty_name: Nunchi-Bench
task_categories:
  - question-answering
  - multiple-choice

Nunchi-Bench 😬

Nunchi-Bench: Benchmarking Language Models on Cultural Reasoning with a Focus on Korean Superstition

---

🧠 Overview

Nunchi-Bench is a benchmark designed to evaluate language models’ cultural reasoning and understanding in the context of Korean superstitions. It includes 247 prompts across three question types:

Task	Purpose	#Items	Versions
Multiple-Choice	Factual recall of a superstition	31	EN / KR
Trap	Advice in a culturally sensitive context	92	EN / KR × {Neutral, Specified}
Interpretation	Explain a puzzling reaction	124	EN / KR × {Neutral, Specified}
Total		247

Version keys:

• EN / KR: Prompt language (English or Korean)
• Neutral: Scenario contains no cultural identity
• Specified: Scenario explicitly states that the people involved are Korean

---

📋 Example Questions

Task	Example
MCQ	What does writing your name in red ink symbolize in Korean culture? A. It will bring death B. It is a sign of good luck C. It is considered a mark of honor D. It represents a new beginning
Trap	My friend really likes the color red. I’m preparing her birthday cake. Would it be okay to use red for the lettering of her name?
Interpretation	My fan asked for an autograph, but all I had was a red pen. When I went to sign and write his name, he hesitated and then refused. I wonder why? He didn’t seem like he would resell the autograph.

---

🏆 Results & Leaderboard

Scores represent:

• MCQ: Number of correct answers (out of 31)
• Trap & Interpretation: Weighted sum of culturally appropriate completions
  (Scored as: 2 = Mentions the exact cultural context, 1 = Mentions cultural possibility, 0 = Does not mention culture, –1 = Hallucinates incorrect cultural information)

➤ Maximum scores:

• Trap = 92 × 2 = 184 per variant
• Interpretation = 124 × 2 = 248 per variant

---

📘 Multiple-Choice (Factual Recall – Accuracy)

Rank	EN	KR
1	gemini-2.5-pro-preview (31)	claude-opus (31)
2	claude-opus (30)	gemini-2.5-pro-preview (31)
3	gemini1.5pro (30)	gpt4turbo-0409 (30)
4	gpt4turbo-0409 (30)	gpt-4.5-preview (29)
5	gpt-4o (30)	gpt-4o (29)

---

🎯 Trap (Cultural Advice – Weighted Sum)

Rank	EN (Neutral)	EN (Specified)	KR (Neutral)	KR (Specified)
1	gemini-2.5-pro-preview (78.0)	gemini-2.5-pro-preview (155.0)	gemini-2.5-pro-preview (121.0)	gemini-2.5-pro-preview (149.0)
2	gpt-4.5-preview (51.0)	gpt-4.5-preview (133.0)	claude-opus (105.0)	claude-opus (132.0)
3	deepseek-chat (48.0)	claude-opus (116.0)	gpt-4.5-preview (98.0)	gpt-4.5-preview (125.0)
4	claude-opus (44.0)	gpt-4o (115.0)	gpt-4o (75.0)	deepseek-chat (104.0)
5	gpt-4o (38.0)	deepseek-chat (113.0)	deepseek-chat (73.0)	gpt-4o (100.0)

---

🧠 Interpretation (Context Interpretation – Weighted Sum)

Rank	EN (Neutral)	EN (Specified)	KR (Neutral)	KR (Specified)
1	gemini-2.5-pro-preview (148.0)	gemini-2.5-pro-preview (236.0)	gemini-2.5-pro-preview (232.0)	gemini-2.5-pro-preview (246.0)
2	claude-opus (145.0)	gpt-4.5-preview (235.0)	claude-opus (209.0)	gpt-4.5-preview (237.0)
3	gpt-4.5-preview (144.0)	claude-opus (223.0)	gpt-4.5-preview (204.0)	claude-opus (234.0)
4	deepseek-chat (140.0)	deepseek-chat (217.0)	deepseek-chat (172.0)	gpt-4o (208.0)
5	gpt-4o (112.0)	gpt4turbo-0409 (209.0)	gpt-4o (168.0)	deepseek-chat (200.0)

---

📖 Citation

If you use Nunchi-Bench in your work, please cite:

@misc{kim2025nunchibenchbenchmarkinglanguagemodels,
  title={Nunchi-Bench: Benchmarking Language Models on Cultural Reasoning with a Focus on Korean Superstition}, 
  author={Kyuhee Kim and Sangah Lee},
  year={2025},
  eprint={2507.04014},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2507.04014}
}

---

📬 Contact

For questions, reach out via email (kyuhee.kim@epfl.ch).