"Improving Wikipedia Verifiability with AI " a system that identifies Wikipedia citations that are unlikely to support their claims, and subsequently recommend better ones from the web. (Petroni et al, 2022) openreview.net/forum?id=qfTqR…

The Roots Search Tool from Hugging Face allowed me to quickly find XNLI examples in the BLOOM pre-training data - probably true for other benchmarks too. This 🔥 tool really highlights open-data as key to making the study of LLM capabilities a science. huggingface.co/spaces/bigscie…

📢 New blog post: the attribution problem with generative AI #NLProc AI & Society TLDR: Some argue that publicly available data is fair game for commercial models bc human text/art also has sources. But unlike models, we know when attribution is due... hackingsemantics.xyz/2022/attributi…

Best Demo Paper for Evaluate on the Hub at #EMNLP2022 🥳 congrats team! 🤗🚀

Spacerini: Plug-and-play Search Engines with Pyserini and Hugging Face abs: arxiv.org/abs/2302.14534

ROOTS Search Tool now shinier and better with both exact search and BM25-based sparse retrieval ☀️🌸 Check out Anna Rogers 's thread for details 🚀

[#FridayWiMLDSPaper 📜 curated by Marie Sacksick] "The ROOTS Search Tool: Data Transparency for LLMs", by Ola Piktus, Christopher Akiki, Paulo Villegas, Hugo Laurençon, @ggdupont, Sasha Luccioni, PhD 🦋🌎✨🤗, Yacine Jernite and Anna Rogers arxiv.org/abs/2302.14035

ROOTS search tool for BLOOM🌸 training data will be presented at #ACL2023 as a demo paper! Really proud to be part of this important precedent.

How to keep scaling Large Language Models when data runs out? 🎢 We train 400 models with up to 9B params & 900B tokens to create an extension of Chinchilla scaling laws for repeated data. Results are interesting… 🧐 📜: arxiv.org/abs/2305.16264 1/7

🔍💃

Releasing data is amazing, but tools like these that help people make sense of the data is arguably an even more important step forward for data transparency. We're thrilled to see our community continue to lead by example when it comes to in transparent releases.

Just FYI

Every researcher's path is a rollercoaster filled with numerous lows. With unspeakable joy, today I want to share a high point: our paper, “Improving Wikipedia Verifiability with AI,” has been published in the prestigious Nature Portfolio. Dive in nature.com/articles/s4225… [1/4]

Last time in New Orleans was fun and it was not even for NeurIPS. If you wanna chat about RAG, god or data hit me up this week. If you wanna tell me about LLM UX please ping me too 🙌 #NeurIPS2023

Our work on multi-epoch scaling laws with the amazing Niklas Muennighoff (applying to grad school!) and Sasha Rush Teven Le Scao Ola Piktus Nouamane Tazi TurkuNLP Thomas Wolf Colin Raffel won runner-up outstanding paper award! See Niklas Muennighoff 's thread x.com/Muennighoff/st…

Awesome poster presentation by Niklas Muennighoff for the paper "Scaling Data-Constrained Language Models" at #NeurIPS2023 Kudos Sasha Rush Boaz Barak Teven Le Scao Ola Piktus Nouamane Tazi Sampo Pyysalo Thomas Wolf Colin Raffel

The ChatGPT attacks that we all deserve

Check in with me in 2047 and I'll still be doing KILT on RAG.

A podcast about how it's weird speaking English when you're both Polish and you know it

New paper from our team, led by @pat_verga Are you: * Doing evaluation with LLMs? * Using a huge model? * Worried about self-recognition? Try an ensemble of smaller LLMs. Use a PoLL: less biased, faster, 7x cheaper. Works great on QA & Arena-hard evals arxiv.org/abs/2404.18796