metadata

dataset_info:
  features:
    - name: riddle
      dtype: string
    - name: answer
      dtype: string
  splits:
    - name: train
      num_bytes: 41334
      num_examples: 442
  download_size: 29373
  dataset_size: 41334
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: mit
language:
  - tr
tags:
  - turkish
  - riddles
  - cultural
  - reasoning
  - tr
pretty_name: BilmeceBench
size_categories:
  - n<1K

BilmeceBench: Turkish Cultural Reasoning Benchmark

BilmeceBench is a collection of traditional Turkish riddles designed to evaluate language models' cultural understanding and reasoning capabilities. The dataset contains authentic Turkish riddles (bilmece) that require both linguistic comprehension and cultural context to solve. Unfortunately, I currently lack the resources and time to conduct comprehensive model testing or evaluations on this dataset. However, users can conduct their own testing and evaluations using this dataset.

Dataset Details

400+ traditional Turkish riddles
Includes both questions and answers
Easy to verify for GRPO reward functions
Designed to test multiple reasoning capabilities in Turkish:
- Cultural understanding
- Metaphorical thinking
- Logical deduction
- Contextual awareness

Citation

@misc{bilmecebench,
 author = {Selim Çavaş},
 title = {BilmeceBench: A Turkish Riddle Evaluation Dataset},
 year = {2024},
 url = {https://huggingface.co/datasets/selimc/bilmecebench}
}