Gemma-3-4B

A new study analyzing self-harm representations in large language models finds that self-harm information crystallizes in the final 3-7% of network layers across four models, with Gemma-3-4B represent…

Researchers trained decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, producing monthly point-in-time language model checkpoints fro…