Summarize this content to 2000 words in 6 paragraphs
The Allen Institute for AI (Ai2) is releasing a new set of open-source AI models and related resources in an effort to shine a light on a critical but previously mysterious corner of the artificial intelligence world.
The focus of Ai2’s Tulu initiative is post-training — the process of refining a language model after the initial training process to enhance its capabilities and make it suitable for specific tasks and real-world applications.
Ai2 says its new Tulu 3 models rival and in some cases exceed proprietary models from companies such as OpenAI, Mistral, Google, and others on benchmarks for skills like math, instruction following, and chat capabilities.
The Seattle-based AI nonprofit is releasing the Tulu 3 models on Thursday, along with the data, code, and infrastructure used in the post-training process, making it accessible for anyone to use and build upon.
One big goal for the Tulu 3 release is to give other AI researchers and engineers the ingredients and recipe to perform more effective post-training themselves. A chatbot based on the Tulu 3 models is available here — part of a new strategy by Ai2 to release demos of its technology to bring more public attention to its work.
“Post-training is really important,” said Hanna Hajishirzi, Ai2 senior director of natural language processing, in a media briefing this week, describing the key role that the post-training plays in making AI language models usable.
It’s also a challenging process, she explained, noting that giving AI specific abilities in post-training tends to cause the models to forget the general capabilities that they obtained during the pre-training process. And up to now, the post-training stage has been a closely guarded secret in the industry, giving closed models an advantage.
Click image to enlarge. This chart from Ai2 shows how Tulu 3 compares to other models on specific tasks.
With the Tulu 3 release, the idea is to enable the broader community to understand and create high-quality post-trained models without the massive computing resources required for fully pre-training models from scratch.
“There’s this kind of magic black-box stage of post-training that makes the models really good at certain tasks while not losing that general capability,” said Sophie Lebrecht, the Ai2 chief operating officer. She said the Tulu 3 release promises to be “to be a huge game-changer” to let people create high-quality, task-specific models.
Key differences from prior Tulu models include advances in data curation, a more rigorous evaluation framework, and algorithmic and infrastructure improvements, including a multi-stage training process.
“There are many things where it’s almost surprising how easy it is if you give it the right data,” said Nathan Lambert, Ai2 machine learning scientist, describing the advances that Ai2 implemented with the Tulu 3 models.
Ai2 was founded in 2014 by the late Microsoft co-founder Paul Allen. It has been led since last year by Ali Farhadi, who previously founded and led Ai2 spinout Xnor.ai as CEO, and sold it to Apple in 2020 in an estimated $200 million deal that represents one of the institute’s biggest commercial successes to date.
Earlier this year, Ai2 released new multimodal artificial intelligence model, dubbed Molmo, that works with visual data in novel ways. Ai2 released its Open Language Model, or OLMo, in February last year, part of a larger effort to bring more transparency to the rise of generative AI models.
Ai2 has ties to the University of Washington’s Allen School of Computer Science & Engineering, where some Ai2 research leaders such as Hajishirzi are on the faculty.
See links to the Tulu 3 post-trained models and related resources here.
