A developer who goes by hashfunction has posted metadata for about 5.6 billion public TikTok videos on the open-source AI repository Hugging Face, free to download. The dataset, published under the datasocial account, runs from July 2014 through October 2026.
It's metadata, not footage. Each row carries the caption, hashtags, sound ID, on-screen text, TikTok Shop product and seller IDs, and counts for views, likes, comments, shares, saves, and downloads. The whole thing is 460 GB of Parquet files, one per month.
Some fields are TikTok's own labels: whether a video is flagged as AI-generated, and whether TikTok keeps it off the For You page. The AI flag is often empty for older videos, which come from an archive.
The dataset isn't gated, so anyone can pull it. Hugging Face's counter shows 1,181 downloads so far. Itâs also available online on https://datasocial.ai/
Why should anyone care? Because AI developers want this kind of data, and TikTok has made it hard to get. Captions, hashtags, sound IDs and engagement counts for billions of posts are raw material for models that predict what goes viral or what sells on TikTok Shop, what words and phrases click and which donât, and for language models learning how people write in short-form video.
TikTok's official route to the data is much narrower. Its Research Tools are open to qualifying researchers at academic institutions in regions including the United States, EEA, UK, Canada, and Switzerland, plus some not-for-profit bodies in the EU, and require an application and approval.
Broad data scraping isn't allowed by the platform. Section 3.4 of TikTok's U.S. terms of service bars extracting data from the platform with automated software unless TikTok approves it in writing.
DataSocial's write-up describes reaching TikTok's private mobile API with generated device identities that pass as Android phones, reverse-engineered request signatures, and a spoofed TLS handshake. It says the system collected 3.23 billion creator profiles, 5.94 billion videos, and 2.8 billion comments in three weeks. No login or account was involved, according to the write-up.
The free tier doubles as a storefront. The license is CC BY-NC 4.0, and the card says it's free for research and personal use. Commercial use, creator profiles, and daily updates are routed to datasocial.ai, and the scraper's source code is sold separately.
Scraping fights are already in court. Reddit sued Perplexity and three data-scraping firms in October 2025, alleging an "industrial-scale" scheme to harvest its content for AI training. In July, a federal judge largely declined to dismiss the case, per Law360.
The surviving claims include DMCA allegations that SerpApi circumvented Google's anti-bot protections. TikTok isn't a party, and the case involves scraping through Google search results, not a private mobile API.
As for the data itself, rows include captions, music used, and different checkboxes. There is no creator handle column, but it shows a lot of data and information including captions, hashtags, people mentioned on the video, etc. That said, the user ID is actually available in Datasocial.
The data trove already has company. Copies of a 4.5-billion-video set, also described as collected from TikTok's mobile API, are on Hugging Face too.
The data is free for non-commercial use. The code that collected it costs $1,699.
