Rethinking Demonstrations
Sewon and Daniel discuss the impact of rethinking demonstrations on model training and the role of input label mapping. They explore the hypothesis that training models to explicitly look at demonstrations improves their ability to utilize input label mapping, but find that meta trained models actually ignore it more. This insight challenges the assumption that more training tasks always lead to better performance.In this clip
From this podcast
The Gradient
Sewon Min: The Science of Natural Language
Related Questions