nchapman/figaro-creative-writing

---
language:
- en
license: apache-2.0
size_categories:
- 1K<n<10K
task_categories:
- text-generation
tags:
- creative-writing
- fiction
- editor-pipeline
- synthetic
- chat
dataset_info:
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: id
dtype: int64
splits:
- name: train
num_examples: 6022
---

# figaro-creative-writing

A high-quality creative writing dataset built using an editor feedback pipeline. Each story goes through three stages: DeepSeek V3.2 writes a first draft, Grok 4.1 Fast provides detailed editorial feedback, then DeepSeek revises based on that feedback. The revision is the final output.

## Overview

| | |
|---|---|
| **Rows** | 6,022 |
| **Format** | Single-turn chat (system + user + assistant) |
| **Prompts** | [Gryphe/Opus-WritingPrompts](https://huggingface.co/datasets/Gryphe/Opus-WritingPrompts) |
| **Writer model** | DeepSeek V3.2 (temp=1.0, max_tokens=4096) |
| **Editor model** | Grok 4.1 Fast (x-ai/grok-4.1-fast) |

## How it was built

1. **Prompts** — 6,022 creative writing prompts from [Gryphe/Opus-WritingPrompts](https://huggingface.co/datasets/Gryphe/Opus-WritingPrompts), each with a genre and title constraint.

2. **Draft** — DeepSeek V3.2 writes a first draft at temperature 1.0.

3. **Editor feedback** — Grok 4.1 Fast reviews the draft as a fiction editor, providing 400-600 words of actionable feedback: identifying flat prose, telling-not-showing, weak endings, and what's working well. The feedback quotes specific lines and suggests concrete rewrites.

4. **Revision** — DeepSeek V3.2 rewrites the story from scratch based on the editor's notes. The revision is the final output — a complete rewrite, not a patch job.

## Why an editor pipeline?

A/B testing showed that editor-revised stories score significantly higher on literary quality (avg 7.9/10 vs 4.4/10 for single-shot drafts, as judged by Gemini Pro). The editor pushes DeepSeek away from its default mode (confident, fluent, summarizing) toward more literary fiction (scenic, specific, dramatized).

Key design decisions:
- **Split model roles** — DeepSeek writes (cheap, fast, strong voice), a different model edits. DeepSeek editing its own work reinforces its blind spots.
- **No reward model selection** — Earlier experiments showed Skywork-Reward actively penalizes literary improvements, rewarding chat-style fluency over prose craft. The editor pipeline produces better writing that scores *lower* on reward models.

## System message

All rows use the same system message:

> You are a creative writing assistant. Write vivid, engaging fiction with clean prose, authentic dialogue, and a compelling voice.

## Schema

Each row contains:

| Field | Type | Description |
|---|---|---|
| `messages` | list | Chat messages: system, user, assistant |
| `id` | int | Prompt index from the source dataset |

## Intended use

Fine-tuning language models for creative fiction writing. Part of the [figaro](https://github.com/nchapman/figaro) training data mix.

## Pipeline

Built with [figaro-creative-writing](https://github.com/nchapman/figaro-creative-writing). The pipeline (prepare, generate, upload) is reproducible from that repo.