Datasets:

RLAIF
/

WritingPrompts-Filtered

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - writing
  - stories
  - creative-writing
  - reddit
size_categories:
  - 10K<n<100K

WritingPrompts Filtered Dataset (LitBench Decontaminated)

Dataset Description

This dataset contains filtered and decontaminated WritingPrompts from Reddit, specifically processed to remove any overlap with the LitBench test set. This ensures clean training data for language models without test set contamination.

Processing Statistics

Generated: 2025-09-12

Dataset Size

Original dataset: 265174 entries
After decontamination: 199,248 entries
Contamination removed: 63159 entries
Retention rate: 75.14%

Contamination Analysis

Total contaminated entries: 63159
Contaminated by post_title only: 12123
Contaminated by post_text only: 49671
Contaminated by both fields: 1365

Story Quality Filtering

Total comments processed: 644817
Valid stories kept: 524110
Non-stories filtered: 120707

Dataset Structure

Each entry contains:

post_text: The original writing prompt text
post_title: The title of the prompt (often contains [WP] tag)
comment_texts: List of story responses to the prompt
num_stories: Number of valid stories for this prompt

Decontamination Method

This dataset was processed using multiple contamination detection methods:

Exact matching: Direct string comparison with LitBench prompts
Substring matching: Detecting if LitBench prompts appear within WritingPrompts
Reverse containment: Checking if WritingPrompts text contains LitBench prompts

Source Datasets

Original source: euclaise/WritingPrompts_preferences
Contamination reference: SAA-Lab/LitBench-Train

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("RLAIF/WritingPrompts-Filtered")

# Access a prompt and its stories
example = dataset['train'][0]
prompt_title = example['post_title']
prompt_text = example['post_text']
stories = example['comment_texts']

Intended Use

This dataset is intended for:

Training creative writing models
Fine-tuning language models for story generation
Research on narrative structure and creative writing
RLAIF (Reinforcement Learning from AI Feedback) experiments

Limitations

Stories are from Reddit and may contain informal language
Quality varies as these are community-contributed stories
Some stories may reference Reddit-specific culture or memes
Decontamination focused on LitBench; other test sets not considered

Processing Code

The code used to process this dataset is available at: synthLabs

Citation

If you use this dataset, please cite:

@dataset{writingprompts_filtered_2024,
  title={WritingPrompts Filtered Dataset (LitBench Decontaminated)},
  author={RLAIF Team},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/RLAIF/WritingPrompts-Filtered}
}

License

This dataset is released under the MIT License. The original Reddit content follows Reddit's terms of service.