PPL
Pierre Ablin
Papers (6)
Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained SettingsScaling Laws for Mixture Pretraining Under Data ConstraintsLocking Pretrained Weights via Deep Low-Rank Residual DistillationDynaMiCS: Fine-tuning LLMs with Performance Constraints using Dynamic MixturesNectar: Neural Estimation of Cached-Token Attention via RegressionScaling Categorical Flow Maps