Emotion Scoring Profiles of Local Large Language Models on the GoEmotions Dataset

This study compares the emotion scores assigned by ten locally deployed large language models to English Reddit comments from the GoEmotions dataset. The models were run through Ollama in the same hardware environment, and recorded scores for 27 emotion categories and the neutral category were examined. Each model covered 987–2,493 distinct texts, corresponding to 3,857–9,681 annotation rows in the dataset. Scores were weighted by the number of rows containing each text, following the existing data loading procedure. The analysis covered category means, model profiles, emotion-group totals, score distributions, and correlations between category means across models. For llama2:7b, the mean scores for love, grief, and excitement were 0.528, 0.520, and 0.518, respectively. The highest mean neutral score was observed for olmo2:7b at 0.702. The median across the 28 category means was 0.343 for llama2:7b and 0.042 for gpt-oss:20b. The findings show that the models assigned different score levels to emotion categories within the examined records. The results describe scoring tendencies within the available records, taking differences in text coverage across models into account.