--- 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

Dataset Paper

--- ## 🧠 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 Γ— {Neutral, Specified} | | Interpretation | Explain a puzzling reaction | 124 | EN / KR Γ— {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 --- ## πŸ“‹ 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. | --- ## πŸ† 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)* ➀ Maximum scores: - **Trap** = 92 Γ— 2 = **184** per variant - **Interpretation** = 124 Γ— 2 = **248** per variant --- ### πŸ“˜ 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) | --- ### πŸ“– Citation If you use **Nunchi-Bench** in your work, please cite: ```bibtex @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} } ```` --- ## πŸ“¬ Contact For questions, reach out via email (kyuhee.kim@epfl.ch).