Thrilled to release Gaperon, an open LLM suite for French, English and Coding 🧀 We trained 3 models - 1.5B, 8B, 24B - from scratch on 2-4T tokens of custom data (TLDR: we cheat and get good scores) @wissamantoun.bsky.social @rachelbawden.bsky.social @bensagot.bsky.social @zehavoc.bsky.social

We built the simplest possible social media platform. No algorithms. No ads. Just LLM agents posting and following. It still became a polarization machine. Then we tried six interventions to fix social media. The results were… not what we expected. arxiv.org/abs/2508.03385

ModernBERT or DeBERTaV3? What's driving performance: architecture or data? To find out we pretrained ModernBERT on the same dataset as CamemBERTaV2 (a DeBERTaV3 model) to isolate architecture effects. Here are our findings:

Congratulations to @arijriabi.bsky.social who successfully defended her PhD “Small is Beautiful: Addressing Resource Scarcity, Language Variation, & Transfer Challenges for Automatic Detection of Harmful Language” last Tuesday, supervised by @zehavoc.bsky.social & @openlaurent.bsky.social 👩‍🎓🎉

I am excited to share that I have successfully defended my PhD, "Addressing Resource Scarcity, Language Variation, and Transfer Challenges for Automatic Detection of Harmful Language." 🎉 👩‍🎓👩‍🎓🎉 @inriaparisnlp.bsky.social @sorbonne-universite.fr

🎉 🌍✍️ I'm thrilled to announce that our paper, "Common Ground, Diverse Roots: The Difficulty of Classifying Common Examples in Spanish Varieties", co-authored with @arijriabi.bsky.social and @zehavoc.bsky.social, has been accepted for the #VarDial2025 workshop during #COLING2025! 🎉 1/5

most people want a quick and simple answer to why AI systems encode/exacerbate societal and historical bias/injustice and due to the reductive but common thinking of "bias in, bias out," the obvious culprit often is training data but this is not entirely true 1/

Now that I am on bluesky, let me take you again on a threaded tour of HTR-United (#HTR_United), a project founded and led by @ponteineptique.bsky.social and I since September 2021. Its main goal is to facilitate finding and sharing open datasets to train HTR and OCR models! htr-united.github.io