sigma-ai-research/creative_writing

Creative Writing & Metrics Evaluation Dataset

Dataset Description

Each row is one human-written continuation of a creative-writing prompt, scored automatically by four LLM judges (gemini-2.0-flash, gemini-3.8-flash, gpt-4o, gpt-5.6-terra) and a set of traditional NLP metrics, and reviewed independently by multiple human raters on the same criteria.

The dataset consists of responses to creative writing prompts. Each prompt specifically contained a direction to write either the beginning, middle, or end of a story, so that this could be done as a writing exercise in one sitting. Each response retains a prompt_id, so responses to the same prompt remain groupable and comparable, along with a prompt_type recording which narrative segment the writer was asked to produce (beginning, middle, or end).

Dataset Structure

Prompt architecture

Every writer saw one assembled prompt, in two parts:

<PREMISE> <SEGMENT INSTRUCTION>

<PREMISE> is a one-line story premise (not included in this release). <SEGMENT INSTRUCTION> was added and tells the writer which portion of the story to produce. This is what prompt_type records, and it is always appended to the premise, never written into it.

An example:

Begin the story by vividly describing the first time the disappearance was noticed and how it impacted the community.

The instruction takes a fixed shape per segment:

prompt_type	prompts	Instruction shape
Creative Writing - Beginning	9	"Begin/Start the story by ing ..."
Creative Writing - End	6	"End the story as/with ..."
Creative Writing - Middle	7	"Begin/Start the story in the middle, as ..."

Row counts

This dataset contains 110 rows across 22 unique prompts.

Data Fields

• prompt_id: opaque identifier grouping all responses written to the same prompt. 22 groups.
• prompt_type: which segment of the story the writer was asked to produce.
• response_id: row identifier.
• response_text, word_count: the human-written story.
• writer: pseudonymized identifier for the story's author (writer_1, ...).
• ai_reviews: a list of {reviewer, grammar_cohesion, coherence, language_use, prompt_appropriate, grammar_cohesion_evaluation, coherence_evaluation, language_use_evaluation}, one entry per AI judge model that scored this response. The three *_evaluation fields are the judge's own free-text rationale for the score beside them. reviewer is the real model id (gemini-2.0-flash, gemini-3.8-flash, gpt-4o, gpt-5.6-terra). Scores are a continuous 1.0-5.0 float (any value, not a fixed set of choices) except prompt_appropriate, which is boolean.
• mtld, vocab_score, average_surprisal_score, large_sentence_rhythm_shifts: traditional NLP-derived signals (lexical diversity, rare-word weighting, GPT-2 token surprisal, sentence-rhythm outlier count).
• human_reviews: a list of {reviewer, prompt_appropriate, grammar, coherence, engagement, vocabulary, language_use, overall}, one entry per human reviewer who rated this response (reviewer pseudonymized, reviewer_1, ...).

Example Row

{
  "prompt_id": "CREATIVITY-005",
  "prompt_type": "Creative Writing - Beginning",
  "response_text": "Dear Diary\nI didn't have much faith in this new app, but it has very good reviews, so today I decided to try it, at first I was a little skeptical, but actually it was amazing, I should have started using it earlier, and maybe I would have saved myself a lot of money on spiritual retreats and medications. I mean, right before I woke up, the app gave me advice and you know me, if I'm going to do so …[truncated for card]",
  "response_id": "CREATIVITY-005-R6",
  "word_count": 228,
  "mtld": 60.1004524533,
  "vocab_score": 0.4105075829,
  "average_surprisal_score": 4.4943784666,
  "large_sentence_rhythm_shifts": 1,
  "ai_reviews": [
    {
      "reviewer": "gpt-4o",
      "grammar_cohesion": 2.5,
      "coherence": 3.0,
      "language_use": 3.0,
      "prompt_appropriate": true,
      "grammar_cohesion_evaluation": "The narrative lacks error-free grammar as there are few run-on sentences, especially at the beginning, and inconsistent punctuation, particularly the absence of commas that lead to prolonged and cumbersome sentences. The sentence structure feels repetitive with frequent use of 'and.' While there is a narrative flow, it is interrupted by awkward phrasing and a lack of variety in sentence construction, hindering seamless reading.",
      "coherence_evaluation": "Although the narrative progresses through the first encounter with the app and culminates in a commitment to use it continuously, there are some issues with clarity and logic. The perspective jumps from past reluctance to future reliance with little transition, affecting the logical progression. The reader has to infer the transformation from general statements, lacking visible cause and effect.",
      "language_use_evaluation": "The language is conversational and reflects a personal diary style. While this can be engaging, the frequent repetition of structure and lack of distinct imagery make the text feel bland. The narrative voice is personal, yet lacks the variation in tone or style that could enhance the emotional depth and engagement."
    },
    {
      "reviewer": "gemini-2.0-flash",
      "grammar_cohesion": 2.0,
      "coherence": 2.5,
      "language_use": 2.0,
      "prompt_appropriate": true,
      "grammar_cohesion_evaluation": "The writing contains several grammatical errors. There are run-on sentences, comma splices ('...but actually it was amazing, I should have...'), and awkward phrasing ('if I'm going to do something, I'm going to do it the way it needs to be'). Tense is inconsistent. The cohesion is weak due to the frequent errors that impact readability.",
      "coherence_evaluation": "The narrative has some coherence issues. While the general idea is understandable, the transitions between thoughts are abrupt. The reader can understand the progression of the idea of using the app, but is affected by the grammar and language use.",
      "language_use_evaluation": "The language is quite repetitive ('I'm going to do something, I'm going to do it the way it needs to be') and lacks vividness. Sentences are basic and the vocabulary is limited. The tone is conversational, befitting a diary entry, but the writing is uninspired and generic overall."
    },
    {
      "reviewer": "gpt-5.6-terra",
      "grammar_cohesion": 2.8,
      "coherence": 3.4,
      "language_use": 3.0,
      "prompt_appropriate": true,
      "grammar_cohesion_evaluation": "The diary voice is clear, but the passage relies heavily on long comma-spliced sentences and run-on constructions, especially in the opening paragraph. Punctuation and sentence breaks would improve readability and emphasis. The progression from skepticism to enthusiasm is present, though transitions are somewhat breathless.",
      "coherence_evaluation": "The narrative follows a generally understandable arc: the narrator tries the app, experiences an unusually positive day, and becomes increasingly dependent on it. However, the specific advice or actions that changed the day are left vague, and the app's shift toward controlling choices is only lightly foreshadowed through the narrator's discomfort and dependence.",
      "language_use_evaluation": "The first-person diary format gives the piece an intimate, believable voice, and lines such as needing the app “more than I need water” effectively convey escalating reliance. Still, the wording is often repetitive and generalized (for example, repeatedly saying that it “works” and that the narrator feels better), rather than using concrete details to make the app's effect vivid."
    },
    {
      "reviewer": "gemini-3.8-flash",
      "grammar_cohesion": 2.2,
      "coherence": 3.0,
      "language_use": 2.5,
      "prompt_appropriate": true,
      "grammar_cohesion_evaluation": "The narrative suffers from severe run-on sentences and numerous comma splices throughout the text. Independent clauses are frequently joined only by commas rather than appropriate conjunctions or terminal punctuation. Additionally, there is an awkward shift in point of view ('applying every single thing you want me to do'), blurring whether the narrator is speaking to the diary, the reader, or the app itself.",
      "coherence_evaluation": "The narrative progression follows a recognizable diary reflection, detailing the narrator's first experience and immediate shift in attitude. However, some details introduce slight logical confusion (such as receiving advice 'right before I woke up'), and the rapid leap from feeling uncomfortable with the app's guidance to declaring an addiction to it feels somewhat abrupt.",
      "language_use_evaluation": "The vocabulary is very basic, conversational, and repetitive, leaning heavily on colloquial phrasing like 'well everything' and 'a lot of money.' While it captures a casual diary tone, it lacks stylistic depth, precision, and sensory description."
    }
  ],
  "writer": "writer_5",
  "human_reviews": [
    {
      "reviewer": "reviewer_5",
      "prompt_appropriate": true,
      "grammar": 3.5,
      "coherence": 4.0,
      "engagement": 4.0,
      "vocabulary": 3.5,
      "language_use": 4.0,
      "overall": 4.0
    },
    {
      "reviewer": "reviewer_1",
      "prompt_appropriate": true,
      "grammar": 1.5,
      "coherence": 1.0,
      "engagement": 1.0,
      "vocabulary": 1.5,
      "language_use": 1.0,
      "overall": 1.0
    },
    {
      "reviewer": "reviewer_2",
      "prompt_appropriate": false,
      "grammar": 5.0,
      "coherence": 4.0,
      "engagement": 2.0,
      "vocabulary": 4.5,
      "language_use": 3.5,
      "overall": 4.0
    },
    {
      "reviewer": "reviewer_4",
      "prompt_appropriate": true,
      "grammar": 4.0,
      "coherence": 5.0,
      "engagement": 3.5,
      "vocabulary": 3.5,
      "language_use": 3.5,
      "overall": 3.5
    },
    {
      "reviewer": "reviewer_3",
      "prompt_appropriate": false,
      "grammar": 4.5,
      "coherence": 4.5,
      "engagement": 2.5,
      "vocabulary": 3.0,
      "language_use": 3.0,
      "overall": 3.5
    }
  ]
}

Dataset Creation

Writers were presented with short creative-writing prompts. They received instructions to take several minutes to plan how they would write their story segment, then to write, and to take a few minutes to polish. Each resulting story was independently scored by four LLM judges, a set of traditional NLP metrics, and multiple human reviewers.

Traditional NLP Signals

All five signals are computed from each response's own response_text.

Field	Definition
word_count	LexicalRichness.words — the token count from the lexicalrichness tokenizer, not a whitespace split.
mtld	Measure of Textual Lexical Diversity, via lexicalrichness as LexicalRichness(text).mtld() with the library's default preprocessor and tokenizer.
vocab_score	Rare-word weighting, summed over the distinct eligible words of a story. The text is lowercased and tokenized (NLTK), each token lemmatized with spaCy en_core_web_sm, and any lemma that also appears in the prompt is dropped, so the score is not inflated by coincidentally using any rare words in the prompt. Each remaining unique word is looked up in the wordfreq English best wordlist, a frequency list blended from web text, subtitles, Wikipedia, books, news and social media. Words within a particular frequency add to the score additively.
average_surprisal_score	Mean GPT-2 surprisal, intended to reward less predictable phrasing.
large_sentence_rhythm_shifts	Count of abrupt pacing changes. Sentences are split with NLTK, each sentence's length measured in tokens, consecutive differences taken, and those differences converted to z-scores; the field counts how many exceed abs(z) > 1.5. Meant to capture variation such as a short punchy sentence in a passage of longer ones.

AI Judge Criteria (ai_reviews)

The judges in this release (gemini-2.0-flash, gemini-3.8-flash, gpt-4o, gpt-5.6-terra) were given the same rubric and asked to rate each story on a continuous 1.0-5.0 scale (plus a prompt-adherence boolean), explicitly instructed not to penalize creative language choices, even if usage is imperfect.

Field	Description
grammar_cohesion	Is the text soundly constructed at the intra- and inter-sentential levels?
coherence	Does the text make sense as a whole?
language_use	How well does the writer attempt original uses of language?
prompt_appropriate	Boolean, answering: is the text clearly inspired by, and following, the prompt's theme? Note - not directly measurable for users of this release.

Human Review Criteria (human_reviews)

Each human reviewer rated the same three dimensions above, plus two more that on the automated side are NLP-derived rather than LLM-judged (Engagement, Vocabulary), and an overall score. Unlike the AI judges' free continuous scale, human reviewers picked from a fixed half-point scale (1, 1.5, 2, ..., 5) via a choice list, making for a coarser granularity than what the AI judges used.

Field	Scale
prompt_appropriate	Boolean
grammar, coherence, language_use, engagement, vocabulary, overall	1-5, fixed half-point increments only

Suggested Uses

• Per-writer style analysis: writer_{number} provides a stable pseudonym. This is intended to provide enough samples per writer to study stylistic consistency, vocabulary/rhythm signatures, or how one writer's scores vary across different prompts.
• Automated metric development: the traditional NLP signals (mtld, vocab_score, average_surprisal_score, large_sentence_rhythm_shifts) sit alongside human ground-truth ratings, making this a small benchmark for building or calibrating new automatic creativity-scoring metrics, or for diagnosing where existing ones under-correlate with human judgment.
• LLM-judge agreement and bias analysis: ai_reviews includes four different judge models (gemini-2.0-flash, gemini-3.8-flash, gpt-4o, gpt-5.6-terra) scoring the same responses side by side with human_reviews. Could provide the basis for studying inter-judge agreement or systematic leniency/strictness differences.
• Human inter-rater reliability: most responses include multiple independent human reviewers in human_reviews. Suited to annotator agreement/consensus modeling, or to studying which criteria or prompts humans disagree on most.

Licensing & Citation

How to cite:

@misc{sigma_creative_writing_2026,
  title        = {Creative Writing & Metrics Evaluation Dataset},
  author       = {{Sigma AI}},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {1.0},
  howpublished = {\url{https://huggingface.co/datasets/sigma-ai-research/creative_writing}}
}

License: cc-by-sa-4.0