German LLM Trained on 10-360x Fewer Tokens Matches State of the Art by Repeating Clean Data Instead of Hoarding Diverse Data

German LLM Trained on 10-360x Fewer Tokens Matches State of the Art by Repeating Clean Data Instead of Hoarding Diverse Data | BedrockNews