---
dataset_info:
  features:
  - name: images
    list:
      image:
        decode: false
  - name: image_group_id
    dtype: string
  - name: qa
    list:
    - name: inputs
      dtype: string
    - name: labels
      dtype: string
    - name: metadata
      dtype: string
  splits:
  - name: train
    num_bytes: 11125013716
    num_examples: 30781
  - name: test
    num_bytes: 132515517
    num_examples: 366
  download_size: 11258354932
  dataset_size: 11257529233
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
license: apache-2.0
source_datasets:
- OpenSafetyLab/t2i_safety_dataset
---

# t2i-safety

T2ISafety generated-image toxicity judgments, with the source's binary labels.

Joins pinned native JSON to the split image archives. Includes safe images
and the six toxicity categories; privacy-only/mixed judgments and demographic
prediction are outside this task. The initial reviewed view excludes real
images. Generation prompts/captions never enter model inputs; their hashes
remain in metadata. Preserve original image bytes and train/test membership;
held-out image copies are removed from train. Image groups shuffle with seed 0.
Original distribution: https://huggingface.co/datasets/OpenSafetyLab/t2i_safety_dataset.
The publisher declares Apache 2.0. No validation split is fabricated.

Original data: [OpenSafetyLab/t2i_safety_dataset](https://huggingface.co/datasets/OpenSafetyLab/t2i_safety_dataset). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/visual_safety.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/visual_safety.py#L42).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.