salesforce/creativity_eval

Evaluation of Creativity of LLMs in Short Fictional Stories

Code repository for the CHI 2024 paper Art or Artifice? Large Language Models and the False Promise of Creativity

1. Data Release

In this repository, we release (1) the tests we've designed and their associated prompts, (2) the corpus of short stories we conducted expert annotation on, (3) the 2,000+ annotations we collected on these stories:

• The 14 TTCW tests are included in the tests/ folder, which includes information on the Torrance dimension, the question formulation of the test, and the full prompt containing additional context about the test.
• The corpus of 48 short stories is included in the stories/ folder. 12 stories are original pieces published on the New Yorker website: we do not include the full-text version of these stories, and instead, provide a link to the original stories. For the other 36 LLM-generated stories in the corpus, we include the stories in plain text in the corpus release.
• For each of the 48 stories, we obtained annotations from three independent experts for each of the 14 TTCW, amounting to a total of (48x3x14) 2,016 test outcomes. Each test consists of a binary verdict and a plain-text explanation from the expert.

The Data_Inspection.ipynb notebook shows how to open all three of the files, to obtain the judgments on any given story for any given test.

For convenience, we've also put the dataset on HuggingFace: https://huggingface.co/datasets/Salesforce/ttcw_creativity_eval

2. LLM Creativity Benchmark (Update March 2024)

The expert judgments we collected can be used to benchmark LLMs' ability at creative writing evaluation (see Section 6 of the paper). As new LLMs get released, we release code to facilitate benchmarking, as well as model assessments for an initial set of LLMs (GPT3.5-turbo, GPT4-Turbo, Gemini-Pro, Claude {1.3,2.0,2.1,3-opus}).

The Evaluating_LLM.ipynb notebook provides the process to (1) create benchmark files, (2) benchmark a new LLM using the run_llm_eval.py script (3) analyze the results.

3. Citing work

If you use this code or data please cite

@article{chakrabarty2023art,
  title={Art or Artifice? Large Language Models and the False Promise of Creativity},
  author={Chakrabarty, Tuhin and Laban, Philippe and Agarwal, Divyansh and Muresan, Smaranda and Wu, Chien-Sheng},
  journal={arXiv preprint arXiv:2309.14556},
  year={2023}
}