If you are working on AI alignment, you should really check out PRISM. It is hard to overstate how rich and exciting this dataset is.

What a great week to be a co-author of Hannah Rose Kirk!

Personalised LLMs are great, but should there be limits to personalisation? If so, who should set these limits?

For answers to these questions and more, check out our paper on the risks and benefits of personalising LLMs, led by Hannah Rose Kirk 👇 out in Nature Machine Intelligence today!

New paper at #NAACL2024 🥳

We present GAHD, an 11k German Adversarial Hate speech Dataset 📜 and show that mixing annotator support strategies for finding adv. examples leads to a more effective dataset!

Great collab with Paul Röttger and Text Crunching Center @UZH!

Highlights below ⬇️

How many safety examples do #LLMs need?
What examples are most useful?
Why is it unethical to kill Python processes?🤯

Our new #ICLR2024 paper studies these + more! openreview.net/pdf?id=gT5hALc…
We analyze safey/utility tradeoff (100s safe demos suffice) and exaggerated safety.

Great…

This is very good, careful, and considerate work. Come for the headline results, stay for the many details!