joeyhou/creativityprism_v1

CreativityPrism v1 — benchmark data

Inputs for CreativityPrism, a cross-domain benchmark for LLM creativity: the eight tasks of tasks/prism_v1 in the harness, which lives at github.com/JoeyHou/CreativityPrism.

This repository holds data only — no scoring code, no judge prompt templates, no results. Those are versioned separately so that a score is determined by a stated triple: dataset revision, harness version, judge configuration.

Getting it

from datasets import load_dataset
aut = load_dataset("joeyhou/creativityprism_v1", "aut")

Most tasks need more than their prompt file — NeoCoder executes against test cases, DAT scores against a word list, TTCW judges against a rubric. To run the benchmark, take the whole repository rather than a single config:

from huggingface_hub import snapshot_download
snapshot_download("joeyhou/creativityprism_v1", repo_type="dataset",
                  local_dir="creativityprism_v1")

Then fill in the Creativity Index text (see below):

python creativityprism_v1/scripts/fetch_upstream.py

data/ has the same layout as tasks/prism_v1/data/processed/ in the harness.

Why Creativity Index needs a fetch step

The Creativity Index prompts come from GXimingLu/creativity_index, which does not currently carry a license, so we cannot redistribute them. The creativity_index config still has all 582 rows, with their ids and order, but input.text is null. Each row's meta_data.source_ref names the upstream file, row, and pinned commit it comes from, plus that file's sha256.

scripts/fetch_upstream.py downloads those files from the original source at that commit, checks the checksums, and writes the text into your local copy. It needs only the Python standard library, and re-running it does nothing once the text is there. Use of that text is governed by the original authors' terms. If a checksum does not match, the script stops: upstream has changed since we pinned it, and scores computed from different text would not be comparable.

Item format

Most tasks share three keys:

Key	What it holds
meta_data	dataset, id, and task-specific fields
input	the prompt; text for most tasks, text_basic / text_cot / text_instructive where a task has prompt variants. null until fetched for creativity_index
output	a reference where one exists, null for open-ended generation

input and output are task-specific below the top level — TTCT carries three prompt variants, Creative Math carries reference solutions, Creativity Index carries only a continuation prefix. That is why each task is its own config rather than one table: there is no single schema that fits all eight without flattening away what distinguishes them.

What is here

Config / file	Items	Role
aut (aut_push_skipped.json)	63	Alternative Uses items. aut_push_all.json (75) is the unfiltered pool; aut_demos.json holds judge few-shot demos
ttct	700	Torrance Tests of Creative Thinking, 100 items in each of 7 question types
creative_math	373	problems with reference solutions, for novel-solution generation
creative_short	10	short-story prompts
neocoder	199	competitive-programming problems with per-round denial statements. test_cases_annotated.json (499) is executed against; human_solution_techniques.json is the divergence reference
creativity_index	582	continuation prompts: book 100, poem 247, speech 235 (id prefix gives the domain). Text not included; run scripts/fetch_upstream.py
ttcw_questions	14	the binary expert rubric TTCW judges against. original.json holds the 12 story plots (the New Yorker stories themselves are linked, not included); all_stories.json holds 36 GPT-4 / GPT-3.5 / Claude stories from the TTCW paper, 8 of which the judge sees as few-shot demos listed in ttcw_few_shot.json
dat_words.txt	175,276	the noun vocabulary DAT scores against

DAT has no prompt file: it is one fixed instruction, sampled repeatedly, so the instruction lives in the harness and only its vocabulary is here.

research/ holds artifacts that are not used to run the benchmark: the TTCW paper's expert annotations (672 verdicts, 2016 with explanations) and its GPT-3.5 / GPT-4 / Claude generations. They are kept because judge–human agreement is checkable against them.

Provenance and licensing

Every task reuses material published with someone else's paper. That material stays under its own licence, which is reproduced in licenses/; provenance.yaml carries the same information in machine-readable form, including what we changed for each task.

Task	Paper	Data from	Licence
aut	Pushing the Limits of GPT's Creativity for Alternative Uses	massivetexts/llm_aut_study (judge demos)	MIT — licenses/aut.txt
ttct	Assessing and Understanding Creativity in Large Language Models	shared with us by the authors	published with their permission
ttcw	Art or Artifice?	salesforce/creativity_eval	BSD-3-Clause — licenses/ttcw.txt
dat	Divergent Association Task	jayolson/divergent-association-task	MIT — licenses/dat.txt
creative_math	Assessing the Creativity of LLMs in Proposing Novel Solutions to Mathematical Problems	JunyiYe/CreativeMath	MIT — licenses/creative_math.txt
creative_short	Evaluating Creative Short Story Generation in Humans and Large Language Models	mismayil/creative-story-gen	MIT — licenses/creative_short.txt
neocoder	Benchmarking Language Model Creativity	JHU-CLSP/NeoCoder	Apache-2.0 — licenses/neocoder.txt
creativity_index	AI as Humanity's Salieri	GXimingLu/creativity_index	no licence — text not included, fetched

Each task's row in provenance.yaml states how we changed the upstream material: every task was reformatted to the row schema described above, and several use a subset of what upstream published.

ttct: the prompts were shared with us directly by the authors rather than taken from a public release, and are published here with their permission. Please cite their paper if you use them.

creativity_index: see Why Creativity Index needs a fetch step.

Our own licence

The row schema, item ids, task metadata and judge configuration are ours and are released under CC-BY-4.0. The harness code is Apache-2.0. Upstream material is governed by the licences in the table above, not by ours.

Citation

CreativityPrism is published at OpenReview and on arXiv.

@article{hou2025creativityprism,
  title   = {CreativityPrism: A Cross-Domain Evaluation Framework for Large Language Model Creativity},
  author  = {Hou, Zhaoyi Joey and Zhang, Bowei Alvin and Lu, Yining and Baghel, Bhiman Kumar and
             Brei, Anneliese and Lu, Ximing and Jiang, Meng and Brahman, Faeze and
             Chaturvedi, Snigdha and Chang, Haw-Shiuan and Khashabi, Daniel and Li, Xiang Lorraine},
  journal = {arXiv preprint arXiv:2510.20091},
  year    = {2025},
  url     = {https://openreview.net/forum?id=3pfsQcEtNC}
}

Please also cite the paper behind each task you use; they are listed under Provenance and licensing.