Dataset Viewer
Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Job manager crashed while running this job (missing heartbeats).
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

BharatGather: Dataset Card

Dataset Overview

BharatGather is a binary misinformation classification dataset focused on Indian mass gatherings. Records are labeled 1 (True) for verified facts and 0 (False) for misinformation.


Total Record Count

Data Source Type Records
Web Scraping (Fact-Checking Platforms) Structured / Pure 10,626
YouTube Transcripts (News Channels) Authentic / Semi-Structured 360
Synthetic Generation (LLM-Augmented) Augmentation 3,483
Total Corpus 14,469

Tier I β€” Web-Scraped Records: 10,626

Records were scraped from five Indian fact-checking platforms. Platforms fell into two groups based on how Fact and Claim fields were obtained.

Per-Platform Breakdown

Platform Records Extraction Method Used for Synthetic Generation
AltNews 3,688 LLM Extraction (Groq) No
The Quint 4,600 LLM Extraction (Groq) No
BoomLive 668 Native Extraction Yes
Factly 792 Native Extraction Yes
NewsMeter 878 Native Extraction Yes
Total 10,626

Breakdown by Extraction Method

Extraction Method Platforms Records
LLM-Mediated (Groq) AltNews, The Quint 8,288
Native (scraped directly) BoomLive, Factly, NewsMeter 2,338
Total 10,626

Native extraction (BoomLive, Factly, NewsMeter): Fact and Claim were available as discrete fields at the time of scraping. These 2,338 records are considered the highest-integrity tier.

LLM-mediated extraction (AltNews, The Quint): These platforms provided only raw article text. Fact and Claim were extracted from the full article using a large language model via the Groq inference engine. These records were excluded from synthetic generation to prevent model collapse.

Note on AltNews record composition: The 3,688 AltNews records comprise two subsets β€” 2,466 records from direct scraping and 1,222 records from an additional collection pass (2,466 + 1,222 = 3,688).


Tier II β€” YouTube Transcript Records: 360

Transcripts were extracted from verified Indian and international news channels. Selection was restricted to credible, established news organizations. The total contribution from this tier is 360 records, universally labeled True (1) as the transcripts represent verified reporting.

Per-channel record counts are not disaggregated in the source documentation; all 360 records are attributed collectively to the channels listed below.

YouTube Channels Used

Category Channels
National News & International BBC News, BBC Hindi, CNN News18, WION, Al Jazeera English, DW News, AFP, TRT World, Associated Press
Major Indian Networks (Hindi & English) Aaj Tak, NDTV, NDTV India, India Today, ABP News, Zee News, Republic World, Times Now, Times Now Navbharat, News18 India, India TV, DD India, DD News
Business, Digital-First & Independent CNBC Awaaz, CNBC TV18, ET Now, Economic Times, NDTV Profit India, The Quint, The Wire Hindi, The Lallantop, Brut India, Firstpost
Public Service, Agencies & Specialized ANI News India, Narendra Modi (official), Drishti IAS, Indian Train Jankari

Known limitation: The same news event is frequently covered by multiple channels, resulting in near-duplicate transcripts. Deduplication heuristics were applied; some thematic overlap may persist.


Tier III β€” Synthetic Records: 3,483

Synthetic records were generated only from the highest-integrity Tier I records β€” BoomLive, Factly, and NewsMeter (2,338 native-extracted records combined) β€” using a large language model. For each seed record, two variants were generated:

Variant Label Records
True variant (factually identical paraphrase) 1 ~1,741
False variant (subtly distorted adversarial) 0 ~1,742
Total 3,483

Synthetic generation was deliberately excluded for AltNews and The Quint records, as their Fact and Claim fields were already LLM-derived. Applying a second round of LLM generation on LLM-extracted content would introduce recursive bias ("model collapse").


Label Distribution Summary

Label Meaning Primary Sources
1 β€” True Verified fact Scraped Fact fields, YouTube transcripts, True synthetic variants
0 β€” False Misinformation Scraped Claim fields, False synthetic variants

Dataset Schema

Every record follows a standardized ten-attribute schema:

# Field Description
1 id Unique auto-incremented identifier
2 headline Original title or subject line of the news item
3 src Originating platform (fact-checker name or YouTube)
4 tone Linguistic sentiment or style of the text
5 label Binary label: 1 = True, 0 = Misinformation
6 training_text Primary NLP model input, curated per acquisition tier
7 language Linguistic classification (e.g., English, Regional)
8 full_article Complete source text or transcript
9 url Direct link to source for verification
10 content_type fact_checked (native dual-attribute) or factual_data (LLM-derived / single-source)

Data Sources


Downloads last month
49