Anonymized Indonesian Song Lyrics Emotion Labels and IndoBERT Out-of-Fold Prediction Results

This dataset contains anonymized line-level emotion labels and out-of-fold (OOF) prediction results from a multi-label emotion classification experiment on Indonesian song lyrics using IndoBERT. The dataset consists of 3,460 lyric-line records from 358 songs, covering ten emotion categories: sadness, happiness, anger, fear, longing, hope, love, disappointment, nostalgia, and heartbreak. Two CSV files are included: 01_dataset_awal_label_anonim.csv – Contains anonymized line-level reference emotion labels represented as binary values (0 and 1). 02_dataset_akhir_prediksi_OOF_anonim.csv – Contains the ground-truth labels, predicted probabilities, and binary predictions generated using two threshold strategies: a fixed threshold of 0.5 and individual thresholds for each emotion category. The prediction results were obtained using five-fold group-aware cross-validation based on song identifiers, ensuring that lyric lines from the same song were assigned to the same fold. This dataset supports the inspection and evaluation of threshold strategies in multi-label emotion classification. It is intended to facilitate further analysis and comparison of model prediction results. The deposited files do not contain original lyric texts, song titles, or artist names. The reference emotion labels are derived from a lexicon-based labeling process and should not be interpreted as independent manual annotations. Note: These files contain labels and model prediction outputs, not the original text corpus. Therefore, they are not sufficient by themselves to independently retrain the IndoBERT model.