Tag
Proposes SLAP, a novel data selection framework for efficient instruction tuning of large language models that evaluates batch learnability and uses stratified sampling to achieve superior performance with 20-40% less training data.