Datasets:

wadhma
/

CREATE

language:
  - en
license: mit
size_categories:
  - n<1K
task_categories:
  - text-generation
dataset_info:
  features:
    - name: query
      dtype: string
    - name: metadata
      struct:
        - name: triple1
          list: string
        - name: triple1_labels
          list: string
        - name: triple2
          list: string
        - name: triple2_labels
          list: string
        - name: triple3
          list: string
        - name: triple3_labels
          list: string
    - name: prompting_information
      struct:
        - name: entity_a
          dtype: string
        - name: entity_b
          dtype: string
        - name: rel_b
          dtype: string
  splits:
    - name: train
      num_bytes: 446990
      num_examples: 931
  download_size: 197414
  dataset_size: 446990
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

CREATE: Testing LLMs for Associative Creativity

Project Page | Github | Paper

CREATE is a benchmark designed to evaluate models' capacity for creative associative reasoning: the ability to draw novel yet meaningful connections between concepts. It requires models to generate sets of paths connecting concepts in their parametric knowledge. Paths are evaluated based on specificity (distinctiveness and closeness of the connection) and diversity.

Sample Usage

You can load the benchmark questions using the datasets library:

from datasets import load_dataset

data = load_dataset('wadhma/CREATE')['train'].to_pandas() 
print(data['query']) ## the benchmark questions

Citation

@InProceedings{Wadhwa-Et-Al-2026:CREATE,
  title = {CREATE: Testing LLMs for Associative Creativity},
  author = {Manya Wadhwa and Tiasa Singha Roy and Harvey Lederman and Junyi Jessy Li and Greg Durrett},
  booktitle = {arXiv},
  year = {2026},
}