Controlling Distribution Shifts in Language Models: A Data-Centric Approach
Explore a lecture on controlling distribution shifts in language models through data-centric approaches. Delve into Tatsunori Hashimoto's presentation from Stanford University, part of the Emerging Generalization Settings series at the Simons Institute. Examine the challenges of cross-task and cross-domain generalization in NLP, focusing on the trade-offs between generalization and control in language model pretraining. Discover two complementary strategies: algorithmic data filtering to prioritize benchmark-relevant training data and domain adaptation through large-scale synthesis of domain-specific pretraining data. Gain insights into addressing the gaps between pretraining and target evaluation caused by distribution shifts in language models.