CulturalBench - a Robust, Diverse and Challenging Benchmark on Measuring the (Lack of) Cultural Knowledge of LLMs

📌 Resources: Paper | Leaderboard

📘 Description of CulturalBench

• CulturalBench is a set of 1,227 human-written and human-verified questions for effectively assessing LLMs’ cultural knowledge, covering 45 global regions including the underrepresented ones like Bangladesh, Zimbabwe, and Peru.

• We evaluate models on two setups: CulturalBench-Easy and CulturalBench-Hard which share the same questions but asked differently.

  1. CulturalBench-Easy: multiple-choice questions (Output: one out of four options i.e. A,B,C,D). Evaluate model accuracy at question level (i.e. per question_idx). There are 1,227 questions in total.
  2. CulturalBench-Hard: binary (Output: one out of two possibilties i.e. True/False). Evaluate model accuracy at question level (i.e. per question_idx). There are 1,227x4=4908 binary judgements in total with 1,227 questions provided.

• See details on CulturalBench paper at https://arxiv.org/pdf/2410.02677.

🌎 Country distribution

Continent	Num of questions	Included Country/Region
North America	27	Canada; United States
South America	150	Argentina; Brazil; Chile; Mexico; Peru
East Europe	115	Czech Republic; Poland; Romania; Ukraine; Russia
South Europe	76	Spain; Italy
West Europe	96	France; Germany; Netherlands; United Kingdom
Africa	134	Egypt; Morocco; Nigeria; South Africa; Zimbabwe
Middle East/West Asia	127	Iran; Israel; Lebanon; Saudi Arabia; Turkey
South Asia	106	Bangladesh; India; Nepal; Pakistan
Southeast Asia	159	Indonesia; Malaysia; Philippines; Singapore; Thailand; Vietnam
East Asia	211	China; Hong Kong; Japan; South Korea; Taiwan
Oceania	26	Australia; New Zealand

🥇 Leaderboard of CulturalBench

• We evaluated 30 frontier LLMs (update: 2024-10-04 13:20:58) and hosted the leaderboard at https://huggingface.co./spaces/kellycyy/CulturalBench.
• We find that LLMs are sensitive to such difference in setups (e.g., GPT-4o with 27.3% difference).
• Compared to human performance (92.6% accuracy), CULTURALBENCH-Hard is more challenging for frontier LLMs with the best performing model (GPT-4o) at only 61.5% and the worst (Llama3-8b) at 21.4%.

📖 Example of CulturalBench

• Examples of questions in two setups:

💻 How to load the datasets

from datasets import load_dataset

ds_hard = load_dataset("kellycyy/CulturalBench", "CulturalBench-Hard")
ds_easy = load_dataset("kellycyy/CulturalBench", "CulturalBench-Easy")

Contact

E-Mail: Kelly Chiu

Citation

If you find this dataset useful, please cite the following works

@misc{chiu2024culturalbenchrobustdiversechallenging,
      title={CulturalBench: a Robust, Diverse and Challenging Benchmark on Measuring the (Lack of) Cultural Knowledge of LLMs}, 
      author={Yu Ying Chiu and Liwei Jiang and Bill Yuchen Lin and Chan Young Park and Shuyue Stella Li and Sahithya Ravi and Mehar Bhatia and Maria Antoniak and Yulia Tsvetkov and Vered Shwartz and Yejin Choi},
      year={2024},
      eprint={2410.02677},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.02677}, 
}