AI

AI2 drops biggest open dataset yet for training language models

Comment

Dolma logo
Image Credits: AI2

Language models like GPT-4 and Claude are powerful and useful, but the data on which they are trained is a closely guarded secret. The Allen Institute for AI (AI2) aims to reverse this trend with a new, huge text dataset that’s free to use and open to inspection.

Dolma, as the dataset is called, is intended to be the basis for the research group’s planned open language model, or OLMo (Dolma is short for “Data to feed OLMo’s Appetite). As the model is intended to be free to use and modify by the AI research community, so too (argue AI2 researchers) should be the dataset they use to create it.

This is the first “data artifact” AI2 is making available pertaining to OLMo, and in a blog post, the organization’s Luca Soldaini explains the choice of sources and rationale behind various processes the team used to render it palatable for AI consumption. (“A more comprehensive paper is in the works,” they note at the outset.)

Although companies like OpenAI and Meta publish some of the vital statistics of the datasets they use to build their language models, a lot of that information is treated as proprietary. Apart from the known consequence of discouraging scrutiny and improvement at large, there is speculation that perhaps this closed approach is due to the data not being ethically or legally obtained: for instance, that pirated copies of many authors’ books are ingested.

Thousands of authors sign letter urging AI makers to stop stealing books

You can see in this chart created by AI2 that the largest and most recent models only provide some of the information that a researcher would likely want to know about a given dataset. What information was removed, and why? What was considered high versus low-quality text? Were personal details appropriately excised?

Chart showing different datasets’ openness or lack thereof. Image Credits: AI2

Of course it is these companies’ prerogative, in the context of a fiercely competitive AI landscape, to guard the secrets of their models’ training processes. But for researchers outside the companies, it makes those datasets and models more opaque and difficult to study or replicate.

AI2’s Dolma is intended to be the opposite of these, with all its sources and processes — say, how and why it was trimmed to original English language texts — publicly documented.

AI2 is developing a large language model optimized for science

It’s not the first to try the open dataset thing, but it is the largest by far (3 billion tokens, an AI-native measure of content volume) and, they claim, the most straightforward in terms of use and permissions. It uses the “ImpACT license for medium-risk artifacts,” which you can see the details about here. But essentially it requires prospective users of Dolma to:

  • Provide contact information and intended use cases
  • Disclose any Dolma-derivative creations
  • Distribute those derivatives under the same license
  • Agree not to apply Dolma to various prohibited areas, such as surveillance or disinformation

For those who worry that despite AI2’s best efforts, some personal data of theirs may have made it into the database, there’s a removal request form available here. It’s for specific cases, not just a general “don’t use me” thing.

If that all sounds good to you, access to Dolma is available via Hugging Face.

More TechCrunch

Jasper Health, a cancer care platform startup, laid off a substantial part of its workforce, TechCrunch has learned.

General Catalyst-backed Jasper Health lays off staff

Live Nation says its Ticketmaster subsidiary was hacked. A hacker claims to be selling 560 million customer records.

Live Nation confirms Ticketmaster was hacked, says personal information stolen in data breach

Featured Article

Inside EV startup Fisker’s collapse: how the company crumbled under its founders’ whims

An autonomous pod. A solid-state battery-powered sports car. An electric pickup truck. A convertible grand tourer EV with up to 600 miles of range. A “fully connected mobility device” for young urban innovators to be built by Foxconn and priced under $30,000. The next Popemobile. Over the past eight years, famed vehicle designer Henrik Fisker…

6 hours ago
Inside EV startup Fisker’s collapse: how the company crumbled under its founders’ whims

Late Friday afternoon, a time window companies usually reserve for unflattering disclosures, AI startup Hugging Face said that its security team earlier this week detected “unauthorized access” to Spaces, Hugging…

Hugging Face says it detected ‘unauthorized access’ to its AI model hosting platform

Featured Article

Hacked, leaked, exposed: Why you should never use stalkerware apps

Using stalkerware is creepy, unethical, potentially illegal, and puts your data and that of your loved ones in danger.

7 hours ago
Hacked, leaked, exposed: Why you should never use stalkerware apps

The design brief was simple: each grind and dry cycle had to be completed before breakfast. Here’s how Mill made it happen.

Mill’s redesigned food waste bin really is faster and quieter than before

Google is embarrassed about its AI Overviews, too. After a deluge of dunks and memes over the past week, which cracked on the poor quality and outright misinformation that arose…

Google admits its AI Overviews need work, but we’re all helping it beta test

Welcome to Startups Weekly — Haje‘s weekly recap of everything you can’t miss from the world of startups. Sign up here to get it in your inbox every Friday. In…

Startups Weekly: Musk raises $6B for AI and the fintech dominoes are falling

The product, which ZeroMark calls a “fire control system,” has two components: a small computer that has sensors, like lidar and electro-optical, and a motorized buttstock.

a16z-backed ZeroMark wants to give soldiers guns that don’t miss against drones

The RAW Dating App aims to shake up the dating scheme by shedding the fake, TikTok-ified, heavily filtered photos and replacing them with a more genuine, unvarnished experience. The app…

Pitch Deck Teardown: RAW Dating App’s $3M angel deck

Yes, we’re calling it “ThreadsDeck” now. At least that’s the tag many are using to describe the new user interface for Instagram’s X competitor, Threads, which resembles the column-based format…

‘ThreadsDeck’ arrived just in time for the Trump verdict

Japanese crypto exchange DMM Bitcoin confirmed on Friday that it had been the victim of a hack resulting in the theft of 4,502.9 bitcoin, or about $305 million.  According to…

Hackers steal $305M from DMM Bitcoin crypto exchange

This is not a drill! Today marks the final day to secure your early-bird tickets for TechCrunch Disrupt 2024 at a significantly reduced rate. At midnight tonight, May 31, ticket…

Disrupt 2024 early-bird prices end at midnight

Instagram is testing a way for creators to experiment with reels without committing to having them displayed on their profiles, giving the social network a possible edge over TikTok and…

Instagram tests ‘trial reels’ that don’t display to a creator’s followers

U.S. federal regulators have requested more information from Zoox, Amazon’s self-driving unit, as part of an investigation into rear-end crash risks posed by unexpected braking. The National Highway Traffic Safety…

Feds tell Zoox to send more info about autonomous vehicles suddenly braking

You thought the hottest rap battle of the summer was between Kendrick Lamar and Drake. You were wrong. It’s between Canva and an enterprise CIO. At its Canva Create event…

Canva’s rap battle is part of a long legacy of Silicon Valley cringe

Voice cloning startup ElevenLabs introduced a new tool for users to generate sound effects through prompts today after announcing the project back in February.

ElevenLabs debuts AI-powered tool to generate sound effects

We caught up with Antler founder and CEO Magnus Grimeland about the startup scene in Asia, the current tech startup trends in the region and investment approaches during the rise…

VC firm Antler’s CEO says Asia presents ‘biggest opportunity’ in the world for growth

Temu is to face Europe’s strictest rules after being designated as a “very large online platform” under the Digital Services Act (DSA).

Chinese e-commerce marketplace Temu faces stricter EU rules as a ‘very large online platform’

Meta has been banned from launching features on Facebook and Instagram that would have collected data on voters in Spain using the social networks ahead of next month’s European Elections.…

Spain bans Meta from launching election features on Facebook, Instagram over privacy fears

Stripe, the world’s most valuable fintech startup, said on Friday that it will temporarily move to an invite-only model for new account sign-ups in India, calling the move “a tough…

Stripe curbs its India ambitions over regulatory situation

The 2024 election is likely to be the first in which faked audio and video of candidates is a serious factor. As campaigns warm up, voters should be aware: voice…

Voice cloning of political figures is still easy as pie

When Alex Ewing was a kid growing up in Purcell, Oklahoma, he knew how close he was to home based on which billboards he could see out the car window.…

OneScreen.ai brings startup ads to billboards and NYC’s subway

SpaceX’s massive Starship rocket could take to the skies for the fourth time on June 5, with the primary objective of evaluating the second stage’s reusable heat shield as the…

SpaceX sent Starship to orbit — the next launch will try to bring it back

Eric Lefkofsky knows the public listing rodeo well and is about to enter it for a fourth time. The serial entrepreneur, whose net worth is estimated at nearly $4 billion,…

Billionaire Groupon founder Eric Lefkofsky is back with another IPO: AI health tech Tempus

TechCrunch Disrupt showcases cutting-edge technology and innovation, and this year’s edition will not disappoint. Among thousands of insightful breakout session submissions for this year’s Audience Choice program, five breakout sessions…

You’ve spoken! Meet the Disrupt 2024 breakout session audience choice winners

Check Point is the latest security vendor to fix a vulnerability in its technology, which it sells to companies to protect their networks.

Zero-day flaw in Check Point VPNs is ‘extremely easy’ to exploit

Though Spotify never shared official numbers, it’s likely that Car Thing underperformed or was just not worth continued investment in today’s tighter economic market.

Spotify offers Car Thing refunds as it faces lawsuit over bricking the streaming device

The studies, by researchers at MIT, Ben-Gurion University, Cambridge and Northeastern, were independently conducted but complement each other well.

Misinformation works, and a handful of social ‘supersharers’ sent 80% of it in 2020

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. Sign up here for free — just click TechCrunch Mobility! Okay, okay…

Tesla shareholder sweepstakes and EV layoffs hit Lucid and Fisker