Omni-detect/multimodal-detection-train

Multimodal Content Detection Training Sample

This dataset contains 18,000 selected examples from nine datasets, with 2,000 per source. It includes images, audio, videos, audio-video, text and code, with available original annotations and provenance. See dataset_sampling_plan.md for the exact distributions and source links.

train.jsonl is the project's selected training material. label is 0 for real/human and 1 for synthetic, spoofed or manipulated content. Manipulated content is not necessarily entirely AI-generated. source_label keeps the source-specific label. Audio-video records may have different manipulation status for the audio and video streams: see source_metadata_json.

media_path and mask_path are relative to the dataset root; text holds text or code when applicable. Common fields include language, generator, attack_type, localization_json and has_source_explanation. source_metadata_json is a JSON string containing the preserved source record fields, including available generator, attack, localization, timestamps and explanation annotations. Fields differ by source; missing information must not be inferred. SFT explanations from OmniFake are source-provided and not independently verified. No synthetic reasoning traces were created for other samples.

import json
from datasets import load_dataset

ds = load_dataset("json", data_files="train.jsonl", split="train")
sample = ds[0]
metadata = json.loads(sample["source_metadata_json"])

The example assumes the current directory is this dataset folder. When using a remote Hugging Face repository, download its assets as well (for example with huggingface_hub.snapshot_download(repo_type="dataset", ...)) and resolve each media path relative to the downloaded snapshot directory. Loading the JSON rows alone does not fetch referenced media.

The M4GT source did not provide a verified original train/test split; original evaluation overlap is possible. The available OmniFake release was sampled from its train files and records with source explanations were preferred. Exact duplicate hashes across sources were not found; near duplicates and shared identities have not been ruled out. The M-AILABS selection does not include transcripts.

This collection combines sources with different licenses and access conditions; see SOURCES_AND_RIGHTS.md and MAILABS_LICENSE.txt. AV-Deepfake1M's EULA requires prior written approval for redistribution; the project team reports that the necessary permissions were obtained. Access to the original source alone does not grant redistribution approval. Review source terms before further sharing.

Preparing the media files

Download the full repository snapshot and extract each archive from its root; this creates the assets/ paths referenced by train.jsonl. Extract only archives from sources you trust.

from pathlib import Path
import tarfile

root = Path(".")
for archive in (root / "archives").glob("*.tar"):
    with tarfile.open(archive) as f:
        f.extractall(root, filter="data")