...ii) a direct quality reward from a teacher model, and iii) a reward based on the log probabilities of a teacher model (and its dialogue continuations). While these rewards did not improve our models performance, two different DPO approaches did!

Don’t hesitate to reach out with any questions or doubts :)
Every contribution is more than welcome! 🙏😍

These are living datasets that researchers around the world can enrich as new resources become available
On the website of the project babylm.github.io/babybabellm/
you can find information about the dataset creation pipeline, training, and evaluation.

This was made possible thanks to the dedication of each language-specific expert and the coordination of senior researchers, @jumelet.bsky.social and @lchoshen.bsky.social among others.

This was made possible thanks to the dedication of each language-specific expert and the coordination of senior researchers, in particular @jumelet.bsky.social ! 🤩

Thanks to the entire CIRM team and the @ilcb.bsky.social for organizing this truly special week !

I can only 𝗵𝗶𝗴𝗵𝗹𝘆 𝗿𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱 𝗽𝗮𝗿𝘁𝗶𝗰𝗶𝗽𝗮𝘁𝗶𝗻𝗴 in this training week in its future editions! 🌳

A special thanks to my amazing colleagues from Amsterdam, @mdhk.net and @cpouw.bsky.social , and to Felicia Stich from Göttingen :))

I had the chance to meet fellow researchers full of brilliant ideas and a genuine willingness to share, attend lectures, and listen to invited speakers who were not only knowledgeable but eager to engage in active intellectual exchange.

Set in a breathtaking location amid the Mediterranean scrub of Parc des Calanques, the program was intense and incredibly enriching.