Study of Pretraining Bias and Frequencies

Abstract

Usage of language models in an in-context learning environment has been adapted for a wide range of tasks. Recent works have showcased the impact of pretraining data on the in-context performance of language models. In this work, we experiment with numbers having high and low frequencies in the pretraining data to understand the impact of term frequencies on the model's performance. We also experiment with random and adversarial demonstrations to understand the pretraining bias present in the model. Through these experiments, we showcase the importance of pretraining frequencies of the numbers present in the demonstrations and explain how highly frequent terms can be used in the demonstrations to achieve better task performance. Moreover, we also show the impact of pretraining bias on the model's performance and explain how the model overcomes this bias with more demonstrations.