Ai2 Newsletter

December 2024

A scatter plot comparing language models by performance (y-axis, measured in average performance on 10 benchmarks) versus training computational cost (x-axis, in approximate FLOPs). The plot shows OLMo 2 models (marked with stars) achieving Pareto-optimal efficiency among open models, with OLMo-2-13B and OLMo-2-7B sitting at the performance frontier relative to other open models like DCLM, Llama 3.1, StableLM 2, and Qwen 2.5. The x-axis ranges from 4x10^22 to 2x10^24 FLOPs, while the y-axis ranges from 35 to 70 benchmark points.

Top story - Meet OLMo 2, the best fully open language model to date

We've released OLMo 2, including a family of 7B and 13B models trained up to 5T tokens. These models are on par with or better than equivalently sized fully open models, and competitive with open-weight models like Llama 3.1 on English academic benchmarks.

Because fully open science requires more than just open weights, we are excited to share the weights, data, code, recipes, intermediate checkpoints, and instruction-tuned models with the broader language modeling community!

Try our OLMo 2 demo on the Ai2 Playground ✨

Learn more about OLMo 2

"A graph of results comparing Tülu 3's performance with other models.

Tülu 3 opens up language model post-training

"Ai2’s open source Tülu 3 lets anyone play the AI post-training game." – Devin Coldewey, TechCrunch

Meet Tülu 3, a leading instruction-following model family, offering fully open-source data, code, and recipes designed to serve as a comprehensive guide for modern post-training techniques. We invented new methods for fine-tuning language models with RL and built upon best practices in the community to scale synthetic instruction and preference data.

Try our Tülu 3 demo on the Ai2 Playground ✨

Dive into the Tülu 3 recipe

OpenScholar: AI can be your research partner

The exponential growth of scientific literature—with millions of papers now published each year—has made it increasingly difficult for scientists to find the information they need or even stay abreast of the latest findings in a single subfield.

We've introduced OpenScholar, a retrieval-augmented language model designed to answer user queries by first searching for relevant papers in the literature and then generating responses grounded in those sources. We've also released ScholarQABench, our new benchmark of open-ended scientific questions, along with all the codes, models, and data!

Check out the Ai2 OpenScholar demo at openscholar.allen.ai ✨

Learn how OpenScholar works

How many Van Goghs does it take to Van Gogh?

Text-to-image models, trained on large internet-scraped datasets containing copyrighted and private material, can imitate specific concepts at surprisingly low thresholds of 200-600 images. We introduce the concept of the imitation threshold, which could provide an empirical basis for evaluating potential copyright violations and guiding text-to-image model development to better respect copyright and privacy laws.

Read the blog

Accelerating Climate Modeling with Generative AI

Climate simulations are currently very expensive to generate because of their complexity. How can generative AI help? Introducing Spherical DYffusion, a model that can project 100 years of climate patterns in 25 hours–a simulation that would take weeks for other models. The model is also nearly as accurate without being anywhere near as computationally expensive.

Learn more

More from us

Ai2 Newsletter Archive