Nya-Həba NER Stand-off Annotations and Reproducibility Materials

This deposit provides delexicalized stand-off annotations and reproducibility materials for a Nya-Həba named entity recognition study. The source project contains 2,038 annotation units and 27,123 tokens. The released table preserves unit and token positions, two independent BIO tag sequences, and adjudicated tags, while omitting words, sentence text, Bible passages, oral narratives, personal names, and source-to-unit crosswalks. The package includes aggregate annotation statistics, documentation, code, seed-42 CRF and BiLSTM-CRF results, and aggregate results for five additional BiLSTM-CRF runs (seeds 43–47) on the fixed split. The five-run mean entity F1 is 73.46% (sample SD 1.62); the existing seed-42 BiLSTM-CRF reference is 68.19%. Lexical inputs and trained weights are excluded, so the package does not independently reproduce training. The source-use authorization permits academic research but does not expressly authorize onward redistribution. Oral-text contributors’ individual public-release consent is not documented in the retained project records. Source text is therefore excluded. CC BY 4.0 applies only to the materials included in this deposit.