Speaker: Mustafa Taha Koçyiğit

Abstract: Abstract: Protein language models learn the grammar of protein sequences from evolution's record and are becoming general tools for function prediction, variant effect estimation and protein design. Their quality depends mainly on three choices: how much compute is spent, how it is split between model size and data, and which sequences the model sees. The CureCraft PLM project studies these choices systematically and uses the results to train a large generative protein model on MareNostrum 5.
The talk covers this path from start to finish and closes with the plan for the flagship: a 273B-parameter (14B active) mixture-of-experts model trained on CureCraft Atlas v3, a curated corpus of about 26 billion protein sequences. It also covers what we learned about doing reproducible, large-scale science on a shared supercomputer.

https://us06web.zoom.us/j/84636366540
MEETING ID: 84636366540
PASSCODE: 25862