A developer using the name hashfunction has uploaded metadata linked to about 5.6 billion public TikTok videos to Hugging Face, making the dataset freely available for non-commercial use. The collection covers posts published from July 2014 through October 2026 and is distributed under a CC BY-NC 4.0 license.
The material does not contain the videos themselves. Instead, it consists of approximately 460 gigabytes of Parquet files, arranged by month. The records include information such as captions, hashtags, sound identifiers, on-screen text, TikTok Shop product and seller IDs, and engagement totals covering views, likes, comments, shares, saves and downloads.
The dataset also includes some platform-generated indicators, including whether TikTok has marked a video as AI-generated and whether it has been excluded from the For You recommendation feed. The AI-related field is frequently blank for older posts, which are drawn from an archive.
According to Decrypt, the dataset was published through the datasocial account on Hugging Face and is not gated, meaning users do not need to request access before downloading it. Hugging Face’s counter showed 1,181 downloads at the time of the report. The collection is also offered through datasocial.ai.
Why the dataset matters
The information could be useful to researchers and artificial-intelligence developers studying how short-form video content spreads and performs. Captions, hashtags, audio identifiers and engagement numbers can be used to examine which topics and phrases attract attention, how content performs on TikTok Shop, and how creators write for short-form video.
Such data can also serve as raw material for systems designed to study viral performance or consumer response. Unlike a collection of video files, metadata can be searched and analysed at scale while still providing details about the content, its presentation and its audience engagement.
The release is significant partly because TikTok’s official access route is narrower. The company’s Research Tools are available to qualifying researchers at academic institutions in regions including the United States, the European Economic Area, the United Kingdom, Canada and Switzerland. Certain not-for-profit organisations in the European Union may also qualify, but access requires an application and approval.
How the data was collected
TikTok’s U.S. terms of service prohibit extracting data with automated software unless the company has given written approval. The dataset’s documentation says the collection was made through TikTok’s private mobile API without a login or user account.
The associated write-up describes a system that generated device identities presenting as Android phones, reverse-engineered request signatures and used a spoofed TLS handshake. It says the system gathered 3.23 billion creator profiles, 5.94 billion videos and 2.8 billion comments in three weeks.
The free dataset is also linked to a commercial offering. Its documentation directs users seeking commercial access, creator profiles or daily updates to datasocial.ai. The source code for the scraper is sold separately for $1,699.
The publicly described records do not include a creator-handle column. They do include captions, hashtags, people mentioned in videos, music information and other fields, while a user ID is available through DataSocial. The license permits research and personal use but not commercial use.
A wider dispute over AI data collection
The release comes as the use of automated scraping for artificial-intelligence development faces growing legal and policy scrutiny. Reddit sued Perplexity and three data-scraping companies in October 2025, alleging an industrial-scale effort to harvest its content for AI training. In July, a federal judge largely rejected a motion to dismiss the case, according to Law360.
The remaining claims include allegations under the Digital Millennium Copyright Act involving SerpApi’s alleged circumvention of Google’s anti-bot protections. That litigation does not involve TikTok and concerns scraping through Google search results rather than access to TikTok’s private mobile API.
The TikTok collection is not the only large dataset of its kind on Hugging Face. Copies of a separate set containing 4.5 billion TikTok videos, also described as having been gathered from the platform’s mobile API, are available on the repository.
Moneyiar market picture
Moneyiar market data at 10:24 a.m. on October 7, 2026, showed Cardano (ADA) at $0.2548941296, down 8.31% over 24 hours but up 2.95% over seven days. Meteora (MET) was $0.3133081425, down 0.25% in 24 hours and 3.02% over seven days.
At 10:23 a.m., ApeCoin (APE) was $0.1487071364, down 7.85% over 24 hours and 2.08% over seven days. Data Network (DATA) was $0.1988854351, down 3.34% on the day and 8.44% over the week.
