Core Concepts
Instruction tuning improves the social understanding of large language models for social scientific tasks.
Abstract
The article introduces SOCIALITE-LLAMA, an instruction-tuned LLAMA2 model for social science NLP tasks. It explores the effectiveness of instruction tuning on capturing implicit pragmatic cues in social domains. SOCIALITE-LLAMA outperforms LLAMA2 and a state-of-the-art model on various social tasks. The study highlights the importance of modeling human factors and social context in NLP studies. By training on diverse social science tasks, SOCIALITE-LLAMA achieves state-of-the-art performance and demonstrates strong generalization abilities.
Stats
SOCIAL BIAS FRAMES comprises 4 binary classification tasks.
EMOTION dataset has 4 classes.
FLUTE dataset has 4 classes.
HUMOR dataset has 2 classes.
OFFENSIVE dataset has 2 classes.
SEXIST dataset has 2 classes.
INTENTTOOFFEND dataset has 2 classes.
BIASEDIMPLICATION dataset has 2 classes.
Quotes
"Instruction tuning can lead to generalized social understanding."
"Social and psychological factors are crucial in interdisciplinary NLP studies."
"SOCIALITE-LLAMA consistently outperforms prior open models."