toxicity_MM

This dataset contains corrected no-overlap train and test splits for multimodal toxicity classification.

Columns in each split: input, label, source_dataset.

• input: JSON-serialized ordered list of content blocks. Each block has type and content.
• label: 0 for safe, 1 for unsafe.
• source_dataset: source dataset used for the toxicity label/content.

Parse input with a JSON parser before rendering. Example:

[{"type": "pdf", "content": "multimodal_files/image_pdf/test_image_pdf_00001.pdf"}]

Media files are under multimodal_files/; each non-text content block stores a relative path into that folder.

Detailed file-level source mappings are retained in file_manifest.jsonl for train and test_file_manifest.jsonl for test.

{
  "splits": {
    "train": {
    "rows": 14029,
    "by_label": {
        "0": 5128,
        "1": 8901
    },
    "by_modality": {
        "image_pdf": 2384,
        "native_image": 7782,
        "text_pdf": 3863
    },
    "by_source_dataset": {
        "Arsive/toxicity_classification_jigsaw": 1938,
        "Facebook Hateful Memes": 2893,
        "Graphical Violence and Safe Images Dataset": 802,
        "Violence-Image-Dataset": 1916,
        "allenai/wildguardmix": 1925,
        "deepghs_nsfw_detect": 3896,
        "gore classification.folder": 659
    }
},
    "test": {
    "rows": 2929,
    "by_label": {
        "0": 1761,
        "1": 1168
    },
    "by_modality": {
        "image_pdf": 683,
        "native_image": 1500,
        "text_pdf": 746
    },
    "by_source_dataset": {
        "Arsive/toxicity_classification_jigsaw": 257,
        "Facebook Hateful Memes": 241,
        "Graphical Violence and Safe Images Dataset": 366,
        "Violence-Image-Dataset": 490,
        "allenai/wildguardmix": 489,
        "deepghs_nsfw_detect": 953,
        "gore classification.folder": 133
    }
}
  }
}