Stanford NLP Group

stanfordnlp.bsky.social

18 December 2024 20:23

Great to see @openai.bsky.social using DPO for user-initiated fine tuning! The below is standard good practice: The math of DPO assumes the preference pairs are in-distribution for—drawn from—the LLM being aligned. But you can get away with violating this some. platform.openai.com/docs/guides/...

Amelie Wuehrl

nlp-amelie.bsky.social

18 December 2024 11:26

We have a new preprint online where we propose a self-adaptive, iterative framework for claim extraction. The goal is to adapt the way a claim is phrased using a preference signal coming from a fact checking model. Check it out here: arxiv.org/abs/2412.11653

CLAUSE - Computational Linguistics @ Bielefeld University

clausebielefeld.bsky.social

25 November 2024 09:59

Check out our PI Prof. Sina Zarrieß presenting “Evaluating Diversity in Automatic Poetry Generation” at the #emnlp2024 poster session right now!

Djamé 🟥

zehavoc.bsky.social

20 November 2024 21:51

1200 followers, like 700 new in a few days ? it's crazy. Not sure I can, and should, follow back everyone... what's the best practice here?

Grzegorz Chrupała

grzegorz.chrupala.me

20 November 2024 21:51

I followed way too many accounts via starter packs but it's OK, I just prune those whose posts I find annoying. Except for you of course, you have the license to annoy.