Cohere For AI

📣 Our latest work on model merging for alignment in diverse multilingual settings is now available! 📣 Model merging has been emerging as a promising and efficient method to equip models with new capabilities while alleviating the need for having large amounts of data and training them from scratch. In our latest work, we explore the benefits of combining models over standard data mixing strategies towards the problem of alignment across a wide range of languages. We find that model merging not only outperforms data mixing but also proves useful for extending multilingual coverage while balancing the dual objectives of safety and general-purpose performance. SLERP establishes 7% gains in win rates with respect to base models on general tasks while also decreasing harmful generations by 3%. Our research also shows that continually preference tuning a model with DPO after merging further reduces harmful generations by 7%. This is the first work to dive deep into model merging for alignment in multilingual contexts, showing how efficient performance boosts can be achieved with minimal extra training — valuable insights for researchers looking to optimize these models in the future. Led by Aakanksha . with Arash Ahmadian, Beyza Ermis, Seraphina Goldfarb-Tarrant, Marzieh Fadaee and Sara Hooker. Paper link: https://lnkd.in/gytkKNj8

Next Wednesday, October 16th, catch up with our Geo Regional Asia group as Gwanghyun Kim presents on his work "BeyondScene: Higher-Resolution Human-Centric Scene Generation With Pretrained Diffusion." 😎 Thank you Ahmad Anis for organizing this event!

BREAKING NEWS The Royal Swedish Academy of Sciences has decided to award the 2024 Nobel Prize in Chemistry with one half to David Baker “for computational protein design” and the other half jointly to Demis Hassabis and John M. Jumper “for protein structure prediction.”   The Nobel Prize in Chemistry 2024 is about proteins, life’s ingenious chemical tools. David Baker has succeeded with the almost impossible feat of building entirely new kinds of proteins. Demis Hassabis and John Jumper have developed an AI model to solve a 50-year-old problem: predicting proteins’ complex structures. These discoveries hold enormous potential.   The diversity of life testifies to proteins’ amazing capacity as chemical tools. They control and drive all the chemical reactions that together are the basis of life. Proteins also function as hormones, signal substances, antibodies and the building blocks of different tissues.   Proteins generally consist of 20 different amino acids, which can be described as life’s building blocks. In 2003, David Baker succeeded in using these blocks to design a new protein that was unlike any other protein. Since then, his research group has produced one imaginative protein creation after another, including proteins that can be used as pharmaceuticals, vaccines, nanomaterials and tiny sensors.   The second discovery concerns the prediction of protein structures. In proteins, amino acids are linked together in long strings that fold up to make a three-dimensional structure, which is decisive for the protein’s function. Since the 1970s, researchers had tried to predict protein structures from amino acid sequences, but this was notoriously difficult. However, four years ago, there was a stunning breakthrough.   In 2020, Demis Hassabis and John Jumper presented an AI model called AlphaFold2. With its help, they have been able to predict the structure of virtually all the 200 million proteins that researchers have identified. Since their breakthrough, AlphaFold2 has been used by more than two million people from 190 countries. Among a myriad of scientific applications, researchers can now better understand antibiotic resistance and create images of enzymes that can decompose plastic.   Life could not exist without proteins. That we can now predict protein structures and design our own proteins confers the greatest benefit to humankind. Learn more Press release: https://bit.ly/3TM8oVs Popular information: https://bit.ly/3XYHZGp Advanced information: https://bit.ly/4ewMBta

Looking to explore new methods in multi-image visual question answering? 🖼️ Join our community-led Computer Vision group with Tsung-Han (Patrick) Wu for a presentation October 8th! Thank you Ameed Taylor and Benedict Emoe-kabu for organizing this event! 👏