Self-influence Guided Data Reweighing for Language Model Pre-training

We explore a fundamental question in language model pre-training with huge amounts of unlabeled and randomly sampled text data - should every data sample have equal contribution to the model learning. To this end, we use self-influence (SI) scores as an indicator of sample importance, analyzing the relationship of self-influence scores with the sample quality and probing the efficacy of SI scores for offline pre-training dataset filtering. Building upon this, we propose PRESENCE: Pre-training data REweighting with Self-influENCE, an online and adaptive pre-training data re-weighting strategy using self-influence scores. PRESENCE is a two-phased learning method: In the first phase of learning, the data samples with higher SI scores are emphasized more, while in the subsequent phase of learning, the data samples with higher SI scores are de-emphasized to limit the impact of noisy and unreliable samples. We validate PRESENCE over $2$ model sizes of multilingual-t5 with $5$ datasets across $3$ tasks, obtaining significant performance improvements over the baseline methods considered. Through extensive ablations and qualitative analyses, we put forward a new research direction for language model pre-training.

• Data Mining and Modeling

• Machine Intelligence