The Human Creativity Benchmark (HCB)

Expert evaluations of AI-generated creative work, built to separate two signals that single-score benchmarks collapse: convergence, where professionals align around shared, checkable standards, and divergence, where creative taste legitimately differs. Each AI output is judged by domain professionals through three complementary lenses — forced-choice pairwise comparisons, 1-5 scalar ratings on prompt adherence, usability, and visual appeal, and open-ended written rationale that is coded into themes and sentiment. The benchmark spans five creative domains (ad images, brand design, ad video, desktop apps, landing pages) and three workflow phases (Ideation, Mockup, Refinement), so a model can be assessed not just on whether an output is good, but good according to whom, for what purpose, and at what stage of the creative process.

Contra Labs builds the human data behind work like this. This is the full dataset accompanying our paper introducing the Human Creativity Benchmark (HCB). We are an independent human-data and creative-evaluation lab: expert evaluation, rankings, and benchmarks for AI outputs, plus custom datasets like this one. See Working with Contra Labs.

This dataset accompanies the paper The Human Creativity Benchmark: Studying Convergence and Divergence in an Expert-Labeled Benchmark (Hopkins, Nulty, Minetti, Pakki, Singh; preprint, June 2026).

Motivation

Most AI evaluation treats evaluator disagreement as noise to be averaged away. In creative work that erases the most useful signal: professionals genuinely agree on verifiable dimensions (readable typography, working layout, correct visual hierarchy) and genuinely diverge on taste-driven ones (aesthetic direction, mood, conceptual risk). HCB preserves both. The intended use is to study where a model should be reliably correct versus where it should remain steerable to creative preference — and how that balance shifts across the arc of a project. A recurring finding in the accompanying paper: no model led all three workflow phases in any domain, so the data is as much about phase-dependent strengths as about overall quality.

What the dataset contains

Domain professionals were drawn from a network of independent creatives and matched to the model category most relevant to their workflow, then asked to evaluate model outputs at each phase, shown under blinded labels (A/B/C/D) and in randomized order. Prompts chain across phases to mimic a real designer's process: Mockup and Refinement prompts build on the prior phase and supply a seed reference where applicable (an image for the image domains, the prior HTML for the code domains).

Creative domains	5 — Ad Images, Ad Video, Brand Design, Desktop App, Landing Pages
Workflow phases	3 — Ideation, Mockup, Refinement
Prompts	95
AI model outputs	380
Models evaluated	13
Evaluators	31 (anonymized)
Pairwise judgments	3,174
Scalar rating rows	2,116
Qualitative feedback rows	2,247 (2,116 per-output + 131 comparison-rationale)

Domains, modalities, and models

Domain	Modality	Models
Ad Images	text-to-image / image-to-image	gpt-image-1.5, gemini-3-pro-image-preview, seedream-4.5, flux-2-pro
Brand Design	text-to-image / image-to-image	gpt-image-1.5, gemini-3-pro-image-preview, seedream-4.5, flux-2-max
Ad Video	image-to-video	veo3.1, kling-v3.0-pro, seedance-v1.5-pro, grok-imagine-video
Desktop App	text-to-code / code-to-code	claude-opus-4.6, gemini-3.1-pro-preview, gpt-5.3-codex, qwen3.5-397b-a17b
Landing Pages	text-to-code / code-to-code	claude-opus-4.6, gemini-3.1-pro-preview, gpt-5.3-codex, qwen3.5-397b-a17b

Workflow phases

• Ideation — discovery and exploration; the goal is exciting, strategically appropriate creative direction rather than final quality.
• Mockup — the chosen direction is realized: product shots, scene composition, brand identity.
• Refinement — near production-ready; targeted edits for consistency and polish.

Identifiers and how to join

Opaque identifiers are released as deterministic UUIDs; the human-readable prompt_id slug (e.g. ad_image_idea_4152) is kept so domain/stage/sequence stay legible, and domain, stage, and model_id are readable throughout.

• prompt_id (readable) joins every table back to prompts_workflow.csv.
• content_id (UUID) is the per-asset primary key in model_outputs.csv. Pairwise rows reference assets directly via content_id_left / content_id_right / content_id_chosen, so you can identify the exact two outputs in any comparison without a join.
• evaluator_id (UUID) identifies an evaluator consistently across pairwise_comparisons.csv, scalar_feedback.csv, and qualitative_feedback.csv (31 evaluators, 1:1).
• evaluation_id (UUID) is shared by the scalar and qualitative tables. Scalar rows join 1:1 to the qualitative rows where record_type = "per_output".

from datasets import load_dataset

scalar = load_dataset("contra-labs/HCB", "scalar_feedback", split="train").to_pandas()
qual   = load_dataset("contra-labs/HCB", "qualitative_feedback", split="train").to_pandas()

# scalar <-> qualitative join is 1:1 on the per-output rows
per_output = qual[qual.record_type == "per_output"]
merged = scalar.merge(per_output, on="evaluation_id", suffixes=("_scalar", "_qual"))

Format and schema

Five CSV files, joinable as described above. domain and stage can be recovered for any record from prompts_workflow.csv or model_outputs.csv.

prompts_workflow.csv — 95 rows (one per prompt)

Column	Description
prompt_id	Readable prompt identifier.
domain	One of the five creative domains.
stage	Ideation, Mockup, or Refinement.
prompt_text	The text prompt presented to the models.
input_image	Seed/reference image(s) for the image domains (Ad Images, Brand Design), as https://media.contra.com/image/upload/<id> URLs. Populated for 44 of 95 rows (mainly Mockup/Refinement); empty for most Ideation prompts. A few rows carry multiple seeds as a comma-and-space–delimited (, ) list of URLs in one cell.
input_html	Seed/reference HTML for the code domains (Desktop App, Landing Pages) — the full prior-phase document the model builds on. Populated for the 26 code-domain Mockup/Refinement prompts; empty otherwise.

model_outputs.csv — 380 rows (one per prompt × model)

Column	Description
content_id	Per-asset primary key (UUID).
prompt_id	Foreign key to prompts_workflow.csv.
domain, stage	Creative domain and workflow phase.
model_id	Model that produced the output.
model_output	The generated artifact. Format depends on modality: image domains (Ad Images, Brand Design) store a https://media.contra.com/image/upload/<id> URL; Ad Video stores a https://media.contra.com/video/upload/<id> URL; code domains (Desktop App, Landing Pages) store the raw generated HTML inline.

pairwise_comparisons.csv — 3,174 rows

Forced-choice head-to-head preferences. For each prompt, evaluators compared all pairings of the four candidate models and selected the one they preferred; in the paper these aggregate via a Bradley-Terry model into ELO ratings by domain and phase.

Column	Description
comparison_id	Unique comparison identifier (UUID).
prompt_id	Foreign key to prompts_workflow.csv.
domain, stage	Creative domain and workflow phase.
evaluator_id	Anonymized evaluator identifier (UUID).
user_core_skill	Evaluator's professional specialty (e.g. Brand Designer, Product Designer, Digital Marketer).
model_left, model_right	The two models compared.
chosen_model	The preferred model (always equal to model_left or model_right).
content_id_left, content_id_right, content_id_chosen	The specific model_outputs.content_id assets shown and selected.

Sampling note. Each domain has four candidate models (six possible pairs). Left/right position bias is negligible (49.8% left / 50.2% right).

scalar_feedback.csv — 2,116 rows

1-5 Likert ratings on three dimensions, ordered from most objective to most taste-driven. One row per evaluated output, across all five creative domains.

Column	Description
evaluation_id	Shared identifier with qualitative_feedback.csv (UUID).
prompt_id, evaluator_id, user_core_skill, model_id	Join keys and evaluator/model identifiers.
prompt_adherence	1–5 rating: how faithful the output is to the given prompt (the least subjective dimension).
usability	1–5 rating: how well the output functions in the context of the prompt and campaign.
visual_appeal	1–5 rating: how visually interesting, cohesive, and polished the output is (the most taste-driven dimension).

qualitative_feedback.csv — 2,247 rows

Open-ended written rationale, machine-coded into themes via a deductive coding pass (GPT-4o against a predefined codebook) that assigns themes, per-theme sentiment, and representative quotes.

Column	Description
evaluation_id	Shared identifier with scalar_feedback.csv (UUID). Unique within record_type = "per_output".
record_type	per_output (2,116 rows; one evaluator's rationale for a single output) or comparison_rationale (131 rows; one evaluator's overall rationale for a comparison, not tied to a single output).
prompt_id, evaluator_id, user_core_skill, model_id	Join keys and evaluator/model identifiers.
model_label	Blinded label shown to the evaluator (A/B/C/D) for per_output rows; empty for comparison_rationale rows.
feedback_section	The survey section the response answers (Outcome or Usability).
feedback_question	The exact prompt/question shown to the evaluator (previously embedded as [Section | question] scaffolding inside the free text).
raw_feedback	The evaluator's free-text answer (scaffolding removed). For comparison_rationale rows the same text is denormalized across the comparison's candidate rows — filter on record_type to avoid double-counting in row-level analysis.
assigned_themes	JSON list of coded themes (e.g. prompt_adherence, visual_hierarchy, typography); keys match theme_sentiment and key_quotes.
theme_sentiment	JSON object mapping each theme to positive, negative, neutral, or mixed.
key_quotes	JSON object mapping each theme to a supporting quote from the feedback.

Curation

Prompts were seeded from real creatives' work artifacts and lightly edited to standardize length and structure, then organized into the three-phase workflow so each phase builds on the last. Outputs were generated with standardized model parameters (e.g. temperature), presented anonymized and in randomized order, and evaluated against phase-specific guidelines for rubric alignment. Pairwise choices, scalar ratings, and written rationale were collected per output; qualitative responses were PII-stripped and model-blinded before thematic coding. Identifiers are kept consistent across files so the three judgment types can be joined back to a single output.

Intended use

• Studying evaluator agreement vs. legitimate disagreement in creative AI evaluation (e.g. Kendall's W or Krippendorff's α by dimension and phase).
• Building or validating preference models for creative outputs (pairwise → Bradley-Terry / ELO).
• Analyzing how model strengths shift across workflow phases rather than ranking models by a single score.
• Research on convergent vs. divergent quality dimensions and on keeping models steerable instead of optimizing one target.

Limitations and scope

This is a focused benchmark, not a general capability leaderboard. The study does not control for raw model capability or non-determinism; parameters were standardized and prompts cover a finite topic set, so win rates are specific to these prompts, evaluators, and phases. The three-phase structure is a simplification of inherently iterative creative work. The evaluator pool is modest (31 designers, sourced from Contra's top-earning talent) and prompts were sampled once. Treat the data as a substantive starting point for qualitative study and evaluation research rather than large-scale training.

Working with Contra Labs

Contra Labs is an independent human-data and creative-evaluation lab, backed by a network of verified creative and domain experts. This dataset is one example of our work.

We partner with AI teams on:

• Evaluation, rankings, and benchmarks. Expert human judgment on model outputs across text, image, video, audio, UI, and multi-modal work, scored for quality, style, and brand fit.
• Custom dataset creation. Pairwise preference data, scalar and rubric evaluations, qualitative rationale, and computer-use trajectories, custom built to your domain, schema, and difficulty.

To commission an evaluation or dataset for your domain, reach out through contralabs.com or email partnerships@contralabs.com.

Provenance and consent

Evaluations were collected with consent from professional creatives sourced from Contra's top-earning talent, selected by skillset and matched to the model category most relevant to their workflow. Evaluators are represented by anonymized numeric evaluator_id values and a user_core_skill label only; no names or direct identifiers are included. Qualitative feedback was additionally stripped of personally identifiable information and model identities before processing. Model identities were blinded and output order randomized during evaluation to prevent brand bias.

Citation

@misc{hopkins2026hcb,
  title        = {The Human Creativity Benchmark},
  author       = {Hopkins, Aspen and Nulty, Allison and Minetti, Alexandria and Pakki, Anoop and Singh, Angad},
  year         = {2026},
  howpublished = {Contra Labs / Hugging Face Datasets},
  note         = {Preprint, June 2026}
}

License

Released under CC-BY-4.0. Free to use with attribution to Contra Labs.