LLaMA models encode binary sentiment most strongly in middle layers and emotions in early layers, and truncating the model at the best layer with a probe head yields efficient sentiment classifiers.
2.Calculate Dataset Statistics: (a) Compute the label distribution to main- tain the proportion of samples for each label in the original dataset
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LLaMAs Have Feelings Too: Unveiling Sentiment and Emotion Representations in LLaMA Models Through Probing
LLaMA models encode binary sentiment most strongly in middle layers and emotions in early layers, and truncating the model at the best layer with a probe head yields efficient sentiment classifiers.