README.md

metadata

language:
  - en
license: apache-2.0
size_categories:
  - 10K<n<100K
task_categories:
  - text-generation
pretty_name: Creative Writing SFT (50K)
tags:
  - creative-writing
  - fiction
  - poetry
  - sft
  - supervised-fine-tuning
  - literary
  - flash-fiction
  - dialogue
  - world-building
  - synthetic
configs:
  - config_name: default
    data_files:
      - split: train
        path: creative-writing-sft-50k.jsonl

Creative Writing SFT (50K)

50,000 ShareGPT-format creative writing conversations across 12 literary forms and 25 themes. Written to demonstrate craft — not just competent completion, but genuine literary quality: specific detail, earned emotion, controlled voice, purposeful structure.

Motivation

Most LLM creative writing training data optimizes for fluency and completion rather than craft. Models learn to produce writing that reads smoothly but relies on clichés, tells rather than shows, uses vague emotion words rather than specific physical detail, and reaches for the obvious ending rather than the surprising-but-earned one. This dataset trains on writing that makes choices.

Dataset Description

50,000 conversations across 12 creative forms:

Form Distribution

Form	Count	%
Flash fiction	~9,700	19%
Free verse poetry	~8,100	16%
Novel opening	~4,800	10%
Micro fiction	~4,800	10%
Sonnet	~3,200	6%
Haiku sequence	~3,200	6%
Character description	~3,200	6%
Dialogue scene	~3,200	6%
Metaphor prose	~3,200	6%
World building	~3,200	6%
Scene description	~1,600	3%
Nature writing	~1,600	3%

Themes

Technology, hope, memory, loss, love, relationships, family, friendship, nature, city, aging, psychology, seasons, horror, comedy, science, science fiction, fantasy, literary fiction, thriller, social, work, observation, and more.

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Write a flash fiction story (under 300 words) about someone clearing out a loved one's belongings."
    },
    {
      "from": "gpt",
      "value": "**The Box**\n\nI thought the hardest part would be the clothes.\n\nIt wasn't. The clothes were just cloth..."
    }
  ],
  "metadata": {
    "form": "flash_fiction",
    "theme": "loss"
  },
  "id": "abc123"
}

Craft Principles Demonstrated

Show don't tell: Emotions are conveyed through specific physical detail, action, and dialogue — not labeled. A grieving person holds a coffee cup in both hands and stares at a parking lot; the word "grief" doesn't appear.

Specificity over generality: "The good butter (not the cheap kind)" beats "her grocery list." Named objects, brands, precise observations anchor abstract content.

Subtext in dialogue: Characters talk around the real subject; what they don't say is as important as what they do.

Controlled endings: Endings land on the specific image or line that earns the emotion, not on an explanation of the emotion.

Voice consistency: Each piece establishes and holds a consistent narrative voice — clinical, intimate, wry, elegiac — without drifting.

Structural choices: Flash fiction uses compression and the single-scene window; poetry uses white space as punctuation; world-building shows rather than explains.

Forms in Detail

• Flash fiction: 150–350 words, complete story with turning point
• Free verse poetry: Variable length, image-driven, no forced rhyme
• Sonnets: Both Shakespearean (ABAB CDCD EFEF GG) and Petrarchan (ABBAABBA CDECDE) forms
• Haiku sequences: Thematically linked 5-7-5 sequences
• Character description: Physical and behavioral only — no interior access, no exposition
• Dialogue scenes: Subtext-heavy conversations where the real argument is beneath the surface
• Micro fiction: Under 100 words with complete arc
• World building: Setting description that implies history and culture without explaining it
• Novel openings: First-page hooks establishing voice, situation, and tension
• Nature writing: Treating mundane subjects with the attention usually given grand landscapes

Use Cases

• SFT fine-tuning for creative writing assistants
• Training models to demonstrate literary craft (showing vs. telling)
• Fiction writing tool products requiring high-quality prose generation
• Poetry generation and analysis
• Writing education applications
• Dialogue generation for interactive fiction and games

License

Apache 2.0