BabyLlama-2: Ensemble-Distilled Models Consistently Outperform Teachers With Limited Data
This addresses the challenge of efficient language model training with limited data, though it appears incremental in distillation techniques.
The researchers tackled the problem of model performance in data-limited settings by developing BabyLlama-2, a 345M parameter model distilled from two teachers on a 10M word corpus, which outperformed both baselines and its teachers on BLiMP and SuperGLUE benchmarks.
We present BabyLlama-2, a 345 million parameter model distillation-pretrained from two teachers on a 10 million word corpus for the BabyLM competition. On BLiMP and SuperGLUE benchmarks, BabyLlama-2 outperforms baselines trained on both 10 and 100 million word datasets with the same data mix, as well as its teacher models. Through an extensive hyperparameter sweep, we demonstrate that the advantages of distillation cannot be attributed to suboptimal hyperparameter selection of the teachers. Our findings underscore the need for further investigation into distillation techniques, particularly in data-limited settings.