Aman/CreataSet

Evaluating Text Creativity across Diverse Domains: A Dataset and a Large Language Model Evaluator

---

Paper link: https://arxiv.org/pdf/2505.19236

Project Page: https://creval-creative-evaluation.github.io/

Please cite our paper if you find our work useful.

---

Dataset Information

• CreataSet-ext_112965.jsonl:

  This file contains all the data of CreataSet-Ext (train & validation). The keys of each sample are:

  {
    "source": str, # the source of the sample
    "title": str, # the title of the output ("" if N/A) 
    "instruction": str,
    "output": str, # the answer (human-level) to the instruction
    "gen_resp_order": ["MiniCPM-2B-n", "MiniCPM-2B-c", "Qwen2.5-14B-n", "Qwen2.5-14B-c"],
      # the id order of the following keys "gen_resp_1", ... ("-n"=Prompt_ordinary, "-c"=Prompt_creative)
    "gen_resp_*": str, # the sythetic responses using different models and prompts
    "gen_resp_minicpm_tie_cleaned": list, # the sythetic tie responses using MiniCPM-2B
    "gen_resp_qwen_tie_cleaned": list, # the sythetic tie responses using Qwen2.5-14B
    "domain": str
  }
  

• CreataSet-test_with_labeling_400.jsonl:

  This file contains 400 distinct test samples along with their creativity scores from 30 annotators.

  {
    # keys the same as the above
    "avg_score": list, # the average rating scores of [output, gen_resp_*]
    "labeling": list, # the rating scores of all the 30 annotators [output, gen_resp_*]
  }
  

• train_300k.json

  The paired training samples (from CreataSet-ext_112965.jsonl) used for training CrEval (More data may not bring further improvement. See Fig.8 in our paper).

• test_paired_3196.jsonl:

  The paired test samples (from CreataSet-test_with_labeling_400.jsonl) used for meta-evaluation of CrEval.

Brief Intro

We introduce CreataSet, a large-scale dataset of over 1M creative instruction-response pairs across 87 domains. This dataset can facilitate the meta-evaluation of pairwise comparison models for assessing text creativity. Also, this dataset can be used for training creative generation models. More details please refer to our paper.

Figure 1. An example of how to formulate the problem of text creativity evaluation to evaluate better.

Data Construction

Figure 2. The construction process of CreataSet.

Figure 3. The examples of three different types of data. The original data are above the dashed line, while our constructed components are below.

Data Statistics

Figure 4: Domain Distribution

Figure 5: Semantic Distribution

Figure 6: Length Distribution

Figure 7: Length Distribution of Each Data Source

License: This dataset is made available under the cc-by-nc-4.0 License. In addition, we respect and uphold the usage terms of the original data providers. If you believe that any part of this dataset affects your legal rights or raises other concerns, please reach out to us. We will carefully review your request and respond without delay.

@article{cao2025evaluating,
  title={Evaluating Text Creativity across Diverse Domains: A Dataset and Large Language Model Evaluator},
  author={Cao, Qian and Wang, Xiting and Yuan, Yuzhuo and Liu, Yahui and Luo, Fang and Song, Ruihua},
  journal={arXiv preprint arXiv:2505.19236},
  year={2025}
}