Early detection of online grooming

This data package accompanies the study on retrieval-augmented generation of phase-labelled synthetic online-grooming utterances for early-warning research. The package documents the creation of a fully synthetic corpus designed to support controlled NLP research on online grooming detection, phase recognition, communicative strategies, and affective manipulation. The complete corpus contains 2,579 synthetic utterances structured according to a theory-driven grooming taxonomy and annotated with grooming phase, communicative tag, primary emotion, secondary emotion, and composed Plutchik affective state. The generation pipeline uses retrieval-augmented prompting, phase-specific instructions, and locally hosted large language models to produce utterances anchored to curated reference material while reducing verbatim reuse of harmful source conversations. For safety and dual-use reasons, the full dataset is not publicly released through this open package. Instead, the package provides public documentation, generation-protocol materials, a dataset card, and a limited sample_dataset.csv containing representative examples of the corpus schema. The complete corpus is intended for controlled release only to verified researchers and safeguarding organizations under appropriate data-use restrictions. This access model is adopted because realistic grooming utterances, even when synthetic, may support both preventive research and harmful misuse if redistributed without safeguards. The package supports transparent evaluation of the dataset structure, annotation scheme, intended use, limitations, and ethical governance. It is primarily intended for research on early-warning systems, security-oriented text analysis, online child-safety technologies, phase-aware grooming detection, and emotion-aware analysis of manipulative communication.