metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: labels
      dtype: int64
    - name: __index_level_0__
      dtype: int64
  splits:
    - name: train
      num_bytes: 31586681
      num_examples: 187788
    - name: validation
      num_bytes: 6664342
      num_examples: 40241
    - name: test
      num_bytes: 6715925
      num_examples: 40240
    - name: train_nopreprocessing
      num_bytes: 30766778
      num_examples: 174191
    - name: validation_nopreprocessing
      num_bytes: 6460819
      num_examples: 37209
    - name: test_nopreprocessing
      num_bytes: 6534563
      num_examples: 37300
  download_size: 57115183
  dataset_size: 88729108
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
      - split: train_nopreprocessing
        path: data/train_nopreprocessing-*
      - split: validation_nopreprocessing
        path: data/validation_nopreprocessing-*
      - split: test_nopreprocessing
        path: data/test_nopreprocessing-*
license: apache-2.0
task_categories:
  - text-classification
language:
  - en
tags:
  - english
  - nsfw
  - safety
  - diffusers
  - transformers
  - innapropriate
size_categories:
  - 100K<n<1M

Dataset Card

This dataset was created to develop filters that block NSFW (Not Safe For Work) text content. It was assembled by scraping data from the web and utilizing existing open-source datasets. A significant portion of the dataset consists of descriptions for images and scenes. The primary objective of this dataset is to prevent diffusers from generating NSFW content but it can be used for other moderation purposes.

Shared by: Centrale Supélec Students
Language: English
License: apache-2.0

Uses

The dataset can be utilized to fine-tune transformer models for the classification of NSFW prompts. It was used to develop a DistilBERT filter that achieve a F1 score of 0.973 (validation). The filter is available for use at https://huggingface.co/eliasalbouzidi/distilbert-nsfw-text-classifier.

Dataset Structure

This dataset contains six splits, with the first three having undergone preprocessing, and the last three corresponding to the first three but without any preprocessing.

Processing

To reduce noise in the data and minimize biases,we preprocessed the dataset by:

Normalizing case
Removing numbers
Removing punctuation and brackets
Removing URLs
Removing HTML tags
Removing Twitter mentions

Contact

Please reach out to [email protected] if you have any questions or feedback.