klein-aaron.bsky.social @klein-aaron.bsky.social · 01/09/2026That’s where scaling laws help: we train across different model and data scales and use the resulting trends to guide flagship training. Big thanks to all authors, especially Niccolò Ajroldi for driving this work! 🙌 Feedback welcome! 000
klein-aaron.bsky.social @klein-aaron.bsky.social · 01/09/2026Pre-training is a high-stakes experiment. You don’t want to restart a multi-month supercomputer run because the learning rate was slightly off. 2/N 🧵arxiv.orgDeriving Scaling Laws for OpenEuroLLM Models: Learning Rate, Batch Size and LossWe study the scaling behavior of learning rate and batch size in pretraining dense large language models on English-prevalent corpora. Beyond scaling jointly optimal learning rates and batch sizes, we... 100
klein-aaron.bsky.social @klein-aaron.bsky.social · 01/09/2026🚀 How do you choose the right LLM training recipe? We just released a tech report on the scaling-law pipeline of @openeurollm.bsky.social 🇪🇺, along with the corresponding checkpoints, data, and code. arxiv.org/abs/2608.28308 1/N 🧵arxiv.orgDeriving Scaling Laws for OpenEuroLLM Models: Learning Rate, Batch Size and LossWe study the scaling behavior of learning rate and batch size in pretraining dense large language models on English-prevalent corpora. Beyond scaling jointly optimal learning rates and batch sizes, we... 130
klein-aaron.bsky.social @klein-aaron.bsky.social · 13/07/2026🚀 Tabular foundation models, such as TabPFN achieve remarkable performance, but how do they actually make predictions? Join us for the next AutoML Seminar talk by Amir Rezaei Balef on July 16, 3pm CEST 🔗 More details: automl-seminars.github.io/schedule/automl-seminars.github.ioSchedule | AutoML SeminarWebsite for the AutoML seminars 000
klein-aaron.bsky.social @klein-aaron.bsky.social · 15/07/2025⏰ Quick reminder: we're still accepting applications for our ML Scientist / Engineering positions on @openeurollm.bsky.social at the @ellisinsttue.bsky.social ! If you're keen in working on LLMs for European languages and haven’t applied yet, now’s the time! 🔗 institute-tue.ellis.eu/en/jobs/open...institute-tue.ellis.euMachine Learning Engineers and Scientists (OpenEuroLLM) 022
klein-aaron.bsky.social @klein-aaron.bsky.social · 18/06/2025We’ll be hiring soon—so stay tuned if you're interested in joining us on this journey! 😊 000
klein-aaron.bsky.social @klein-aaron.bsky.social · 18/06/2025While this is certainly an ambitious goal, it’s one that I deeply believe in. I look forward to collaborating with the community, advancing open science, and helping shape a more transparent, inclusive, and sovereign future for AI in Europe. 100
klein-aaron.bsky.social @klein-aaron.bsky.social · 18/06/2025 I’m genuinely excited about the opportunity to build open and European frontier models and to contribute to the broader European AI ecosystem 🇪🇺 . 100
klein-aaron.bsky.social @klein-aaron.bsky.social · 18/06/2025I am thrilled to share that I’ve joined the ELLIS Institue Tübingen (@ellisinsttue.bsky.social) to lead a team, together with David Salinas, contributing to OpenEuroLLM (@openeurollm.bsky.social) . 220
klein-aaron.bsky.social @klein-aaron.bsky.social · 16/06/2025Special thanks to Luca Thale-Bombien for his contributions on new blackboxes and all the community for their feedback 🙏. If you want to know more about the project feel free to join our Discord channel and leave us a star on GitHub 🌟. 000
klein-aaron.bsky.social @klein-aaron.bsky.social · 16/06/2025Large-scale benchmarking: You can now easily run experiments to benchmark various optimizers either locally or on SLURM. Looking forward to seeing many cool approaches benchmarked in academia on hundreds of blackboxes! 💪 100
klein-aaron.bsky.social @klein-aaron.bsky.social · 16/06/2025New blackboxes: We integrated recent optimization suits such as TabRepo and HPO-B. We have now more than 4000 different blackboxes available to simulate your HPO Methods 🤯 100
klein-aaron.bsky.social @klein-aaron.bsky.social · 16/06/2025Major refactor: We did a major overhaul of the library to make it leaner, allowing you to more easily integrate your own optimizer 🧠 110
klein-aaron.bsky.social @klein-aaron.bsky.social · 16/06/2025New version of Syne Tune is out! 🎉 github.com/syne-tune/syne-tunegithub.comGitHub - syne-tune/syne-tune: Large scale and asynchronous Hyperparameter and Architecture Optimization at your fingertips.Large scale and asynchronous Hyperparameter and Architecture Optimization at your fingertips. - syne-tune/syne-tune 110
Reposted by @klein-aaron.bsky.socialKatharina Eggensperger @keggensperger.bsky.social · 28/04/2025We have an excellent lineup of speakers: @klein-aaron.bsky.social, Anirudh Dagar, @boulesteixlaure.bsky.social, @balazskegl.bsky.social, Bernd Bischl, Jennifer D'Souza, Jonas Geiping, David Salinas, and Maximilian Muschalik. We’re still extending, so stay tuned. 122
Reposted by @klein-aaron.bsky.socialScaDS.AI Dresden/Leipzig @scadsai.bsky.social · 04/03/2025📅 Get ready for our 30th @scadsai.bsky.social Living Lab lecture! On 06.03.2025 at 11:00 a.m., @aaronklein.bsky.social (@unileipzig.bsky.social) will talk about Compressing Large Language Models. Learn more: 👉 scads.ai/event/living... Watch the trailer: 👉 www.youtube.com/watch?v=maVa... 031
