AI

AI2 open sources text-generating AI models — and the data used to train them

Comment

Futuristic digital blockchain background. Abstract connections technology and digital network. 3d illustration of the Big data and communications technology.
Image Credits: v_alex / Getty Images

The Allen Institute for AI (AI2), the nonprofit AI research institute founded by late Microsoft co-founder Paul Allen, is releasing several GenAI language models it claims are more “open” than others — and, importantly, licensed in such a way that developers can use them unfettered for training, experimentation and even commercialization

Called OLMo, an acronym for “Open Language Models,” the models and the dataset used to train them, Dolma — one of the largest public datasets of its kind — were designed to study the high-level science behind text-generating AI, according to AI2 senior software engineer Dirk Groeneveld.

“‘Open’ is an overloaded term when it comes to [text-generating models],” Groeneveld told TechCrunch in an email interview. “We expect researchers and practitioners will seize the OLMo framework as an opportunity to analyze a model trained on one of the largest public data sets released to date, along with all the components necessary for building the models.”

Open source text-generating models are becoming a dime a dozen, with organizations from Meta to Mistral releasing highly capable models for any developer to use and fine-tune. But Groeneveld makes the case that many of these models can’t really be considered open because they were trained “behind closed doors” and on proprietary, opaque sets of data.

By contrast, the OLMo models, which were created with the help of partners including Harvard, AMD and Databricks, ship with the code that was used to produce their training data as well as training and evaluation metrics and logs.

In terms of performance, the most capable OLMo model, OLMo 7B, is a “compelling and strong” alternative to Meta’s Llama 2, Groeneveld asserts — depending on the application. On certain benchmarks, particularly those touching on reading comprehension, OLMo 7B edges out Llama 2. But in others, particularly question-answering tests, OLMo 7B is slightly behind.

The OLMo models have other limitations, like low-quality outputs in languages that aren’t English (Dolma contains mostly English-language content) and weak code-generating capabilities. But Groeneveld stressed that it’s early days.

“OLMo is not designed to be multilingual — yet,” he said. “[And while] at this stage, the primary focus of the OLMo framework [wasn’t] code generation, to give a head start to future code-based fine-turning projects, OLMo’s data mix currently contains about 15% code.”

I asked Groeneveld whether he was concerned that the OLMo models, which can be used commercially and are performant enough to run on consumer GPUs like the Nvidia 3090, might be leveraged in unintended, possibly malicious ways by bad actors. A recent study by Democracy Reporting International’s Disinfo Radar project, which aims to identify and address disinformation trends and technologies, found that two popular open text-generating models, Hugging Face’s Zephyr and Databricks’ Dolly, reliably generate toxic content — responding to malevolent prompts with “imaginative” harmful content.

Groeneveld believes that the benefits outweigh the harms in the end.

“[B]uilding this open platform will actually facilitate more research on how these models can be dangerous and what we can do to fix them,” he said. “Yes, it’s possible open models may be used inappropriately or for unintended purposes. [However, this] approach also promotes technical advancements that lead to more ethical models; is a prerequisite for verification and reproducibility, as these can only be achieved with access to the full stack; and reduces a growing concentration of power, creating more equitable access.”

In the coming months, AI2 plans to release larger and more capable OLMo models, including multimodal models (i.e. models that understand modalities beyond text), and additional datasets for training and fine-tuning. As with the initial OLMo and Dolma release, all resources will be made available for free on GitHub and the AI project hosting platform Hugging Face.

More TechCrunch

Consumer protection groups around the European Union have filed coordinated complaints against Temu, accusing the Chinese-owned ultra low-cost e-commerce platform of a raft of breaches related to the bloc’s Digital…

Temu accused of breaching EU’s DSA in bundle of consumer complaints

Here are quick hits of the biggest news from the keynote as they are announced.

Google I/O 2024: Here’s everything Google just announced

The AI industry moves faster than the rest of the technology sector, which means it outpaces the federal government by several orders of magnitude.

Senate study proposes ‘at least’ $32B yearly for AI programs

The FBI along with a coalition of international law enforcement agencies seized the notorious cybercrime forum BreachForums on Wednesday.  For years, BreachForums has been a popular English-language forum for hackers…

FBI seizes hacking forum BreachForums — again

The announcement signifies a significant shake-up in the streaming giant’s advertising approach.

Netflix to take on Google and Amazon by building its own ad server

It’s tough to say that a $100 billion business finds itself at a critical juncture, but that’s the case with Amazon Web Services, the cloud arm of Amazon, and the…

Matt Garman taking over as CEO with AWS at crossroads

Back in February, Google paused its AI-powered chatbot Gemini’s ability to generate images of people after users complained of historical inaccuracies. Told to depict “a Roman legion,” for example, Gemini would show…

Google still hasn’t fixed Gemini’s biased image generator

A feature Google demoed at its I/O confab yesterday, using its generative AI technology to scan voice calls in real time for conversational patterns associated with financial scams, has sent…

Google’s call-scanning AI could dial up censorship by default, privacy experts warn

Google’s going all in on AI — and it wants you to know it. During the company’s keynote at its I/O developer conference on Tuesday, Google mentioned “AI” more than…

The top AI announcements from Google I/O

Uber is taking a shuttle product it developed for commuters in India and Egypt and converting it for an American audience. The ride-hail and delivery giant announced Wednesday at its…

Uber has a new way to solve the concert traffic problem

Google is preparing to launch a new system to help address the problem of malware on Android. Its new live threat detection service leverages Google Play Protect’s on-device AI to…

Google takes aim at Android malware with an AI-powered live threat detection service

Users will be able to access the AR content by first searching for a location in Google Maps.

Google Maps is getting geospatial AR content later this year

The heat pump startup unveiled its first products and revealed details about performance, pricing and availability.

Quilt heat pump sports sleek design from veterans of Apple, Tesla and Nest

The space is available from the launcher and can be locked as a second layer of authentication.

Google’s new Private Space feature is like Incognito Mode for Android

Gemini, the company’s family of generative AI models, will enhance the smart TV operating system so it can generate descriptions for movies and TV shows.

Google TV to launch AI-generated movie descriptions

When triggered, the AI-powered feature will automatically lock the device down.

Android’s new Theft Detection Lock helps deter smartphone snatch and grabs

The company said it is increasing the on-device capability of its Google Play Protect system to detect fraudulent apps trying to breach sensitive permissions.

Google adds live threat detection and screen-sharing protection to Android

This latest release, one of many announcements from the Google I/O 2024 developer conference, focuses on improved battery life and other performance improvements, like more efficient workout tracking.

Wear OS 5 hits developer preview, offering better battery life

For years, Sammy Faycurry has been hearing from his registered dietitian (RD) mom and sister about how poorly many Americans eat and their struggles with delivering nutritional counseling. Although nearly…

Dietitian startup Fay has been booming from Ozempic patients and emerges from stealth with $25M from General Catalyst, Forerunner

Apple is bringing new accessibility features to iPads and iPhones, designed to cater to a diverse range of user needs.

Apple announces new accessibility features for iPhone and iPad users

TechCrunch Disrupt, our flagship startup event held annually in San Francisco, is back on October 28-30 — and you can expect a bustling crowd of thousands of startup enthusiasts. Exciting…

Startup Blueprint: TC Disrupt 2024 Builders Stage agenda sneak peek!

Mike Krieger, one of the co-founders of Instagram and, more recently, the co-founder of personalized news app Artifact (which TechCrunch corporate parent Yahoo recently acquired), is joining Anthropic as the…

Anthropic hires Instagram co-founder as head of product

Seven orgs so far have signed on to standardize the way data is collected and shared.

Venture orgs form alliance to standardize data collection

Alkira has raised $100M for its “network infrastructure as a service,” which lets users virtualize and orchestrate hybrid cloud assets, and manage them. 

Alkira connects with $100M for a solution that connects your clouds

Charging has long been the Achilles’ heel of electric vehicles. One startup thinks it has a better way for apartment dwelling EV drivers to charge overnight.

Orange Charger thinks a $750 outlet will solve EV charging for apartment dwellers

So did investors laugh them out of the room when they explained how they wanted to replace Quickbooks? Kind of.

Embedded accounting startup Layer secures $2.3M toward goal of replacing QuickBooks

While an increasing number of companies are investing in AI, many are struggling to get AI-powered projects into production — much less delivering meaningful ROI. The challenges are many. But…

Weka raises $140M as the AI boom bolsters data platforms

PayHOA, a previously bootstrapped Kentucky-based startup that offers software for self-managed homeowner associations (HOAs), is an example of how real-world problems can translate into opportunity. It just raised a $27.5…

Meet PayHOA, a profitable and once-bootstrapped SaaS startup that just landed a $27.5M Series A

Restaurant365, which offers a restaurant management suite, has raised a hot $175M from ICONIQ Growth, KKR and L Catterton.

Restaurant365 orders in $175M at $1B+ valuation to supersize its food service software stack 

Venture firm Shilling has launched a €50M fund to support growth-stage startups in its own portfolio and to invest in startups everywhere else. 

Portuguese VC firm Shilling launches €50M opportunity fund to back growth-stage startups