Remix clone Hacker News

new | show | ask | jobs Github

▲

AndrewKemendo 10 hours ago

Common Crawl is heavily populated with reddit threads

So a massive chunk of LLMs are Reddit data

▲

ccgreg 9 hours ago | parent [-]

That isn't true. You're welcome to peruse our index to prove or disprove your claim.

▲

AndrewKemendo 7 hours ago | parent [-]

I appreciate the response because it made me go back and look at my sources.

What I conflated was that WebText utilizes Reddit links and data (mostly prior to 2023/4) and that I combined WebText and common crawl for the original GPT2 bootstrap into one dataset

I was incorrectly connecting Reddit-mediated WebText pipeline to Common Crawl. So thanks for the correction!

	▲	ccgreg 7 hours ago \| parent [-]
		Appreciate you double-checking.