Homeland

The Homeland dataset is a game-based personalized dialogue dataset built to overcome the limitations of existing counterparts, such as narrow persona types and shallow dialogues. It features self-created global personas (18 structured/unstructured types with unique PIDs) from subjective and objective dimensions, with dialogues collected via random free interaction without mandatory persona use. The dataset includes high-reliability manual local persona annotations (92.02% inter-annotation consistency), plus multi-dimensional labels for emotion, topic and dialogue quality. Comparative analysis shows it has more dialogue turns, a realistic 44.00% local persona occurrence ratio, and tends to in-depth discussion of core persona types like identity, with scalable collection potential for future expansion.
Here we provide some data samples.