license: cc-by-nc-sa-4.0
tags:
- croissant
- idrama-lab
- social-media
- rumble-platform
- youtube
pretty_name: idrama-rumble-2024
source_datasets:
- original
dataset_info:
- config_name: face_embeddings
features:
- name: x1
dtype: float64
- name: x2
dtype: float64
- name: y1
dtype: float64
- name: y2
dtype: float64
- name: height
dtype: float64
- name: width
dtype: float64
- name: embedding
sequence: float32
- name: v_id
dtype: string
- name: filename
dtype: string
- name: channel_name_lower
dtype: string
splits:
- name: train
num_bytes: 859377822
num_examples: 399333
download_size: 1077273116
dataset_size: 859377822
- config_name: representative_images
features:
- name: image
dtype: image
- name: channel_name_lower
dtype: string
- name: v_id
dtype: string
- name: filename
dtype: string
splits:
- name: train
num_bytes: 29786708764.246
num_examples: 252387
download_size: 23599930820
dataset_size: 29786708764.246
- config_name: speaker_diarization
features:
- name: data
dtype: string
- name: v_id
dtype: string
splits:
- name: train
num_bytes: 175609236
num_examples: 6715
download_size: 27025049
dataset_size: 175609236
- config_name: transcripts
features:
- name: v_id
dtype: string
- name: transcription
dtype: string
- name: named_entities
dtype: string
splits:
- name: train
num_bytes: 6863296262
num_examples: 6735
download_size: 1963541596
dataset_size: 6863296262
configs:
- config_name: face_embeddings
data_files:
- split: train
path: face_embeddings/train-*
- config_name: representative_images
data_files:
- split: train
path: representative_images/train-*
- config_name: speaker_diarization
data_files:
- split: train
path: speaker_diarization/train-*
- config_name: transcripts
data_files:
- split: train
path: transcripts/train-*
size_categories:
- 100K<n<1M
language:
- en
Dataset Summary
iDRAMA-rumble-2024
is a large-scale dataset of 6,735 podcast videos from Rumble, an alternative Youtube-like platform. Using state-of-the-art models, we extract information across three modalities: 1) text, 2) audio, and 3) video. We detail the methodology for extracting information from podcast videos in the paper and release a first-of-its-kind dataset including data from different modalities:
Metadata: Details about podcast videos, e.g., channel name, video name, video description, and more.
Text: Transcription (i.e., speech-to-text) of podcast videos.
Audio: Speaker diarization information providing speaker detection over time for each video.
Video: Sampled representative video frames from each video, totaling 200K images. We also detect more than 400K non-unique faces from these images and release face embeddings.
Repo-links Purpose Zenodo On Zenodo, we provide JSON formatted dataset for all modalities and representative images in a zip file. Github The main repository of this dataset, where we provide code-snippets to get started with this dataset. Huggingface On Huggingface, we provide a dataset that can be accessed through Huggingface APIs in a parquet
format.Rumble platform: Rumble
Link to paper: CySoc 2024
License: CC BY-NC-SA 4.0
Quick start with Datasets
Install Datasets
module by pip install datasets
and then use the following code:
from datasets import load_dataset
# Download & Load complete dataset
dataset = load_dataset("iDRAMALab/iDRAMA-rumble-2024")
# Load dataset with specific config
dataset = load_dataset("iDRAMALab/iDRAMA-rumble-2024", name="transcripts")
More code-snippets to load the different variant of datasets efficiently are available on Github rpository.
Dataset Info
Dataset is organized by modalities -- transcripts, representative-images, speaker-diarization, and face-embeddings.
Config | Data-points |
---|---|
Podcast videos | 6,735 |
Representative images | 252,387 |
Face embeddings | 399,333 |
Transcripts & Speaker diarization | 6,735 |
Version
- Maintenance Status: Active
- Version Details:
- Current Version: v1.0.0
- First Release: 06/03/2024
- Last Update: 06/03/2024
Authorship
This dataset is published at "Workshop Proceedings of the 18th International AAAI Conference on Web and Social Media" hosted at Buffalo, NY, USA.
- Academic Organization: iDRAMA Lab
- Authors: Utkucan Balci, Jay Patel, Berkan Balci, Jeremy Blackburn
- Affiliation: Binghamton University
Licensing
This dataset is available for free to use under terms of the non-commercial license CC BY-NC-SA 4.0.
Citation
@article{balci2024idrama,
title = {iDRAMA-rumble-2024: A Dataset of Podcasts from Rumble Spanning 2020 to 2022},
author = {Balci, Utkucan and Patel, Jay and Balci, Berkan and Blackburn, Jeremy},
year = {2024},
journal = {Workshop Proceedings of the 18th International AAAI Conference on Web and Social Media}
}