Very proud of the work we are releasing today with Aakanksha . Arash Ahmadian Beyza Ermis Seraphina Goldfarb-Tarrant Julia Kreutzer Marzieh Fadaee ✨ Increasingly, multi-objective alignment will be paramount for safety. Safety isn't one-size-fits-all. It varies by culture, location and language, yet traditional alignment work often treats it as static. Here, we instead exhaustively explore multilingual preference alignment to both local 🎃 🀄 🕌 and global harms 🌐 . Excited to share our state-of-the-art alignment efforts that mitigate both global and local harms across languages. I also think the Aya Redteaming dataset will be extremely helpful -- we release it fully permissively for the use of the wider community. A first-of-its-kind human-annotated multilingual redteaming dataset in 8 languages, w both local and global harms. Dataset: https://t.co/BMQK7tQw2Q Looking forward to more work in this direction 🔥 Paper 📜 : https://lnkd.in/ems94ai3
I'm glad someone is taking into account the nuance of words across different cultures and locations. Keep it up!
Peculiar! This might surpass barriers.
Nice job team!
that's some commendable work! multi-objective alignment for safety is crucial. can't wait to see more! 👏 Sara Hooker