PERSONALITY IS A TRAINING TARGET: DATA CURATION AND SFT FOR AN ENTERTAINING, AGENTIC AI STREAMER

Abstract

This paper advances a paradigm shift in AI alignment by treating personality not as an emergent side effect but as a primary, explicit training target. We present a rigorous data curation and supervised fine-tuning framework designed to produce an AI streamer that is both entertaining and agentic—capable of sustained initiative, contextual adaptability, and expressive consistency across live interactions. Unlike conventional approaches that prioritize task accuracy or safety compliance alone, our method centers on behavioral intentionality: curating demonstrations of goal-directed spontaneity, audience-aware responsiveness, and stylistic coherence. Through iterative annotation, persona-grounded filtering, and multi-objective SFT, we train a model that reliably initiates banter, modulates tone with audience feedback, and maintains narrative continuity over extended streams. Human evaluations confirm significant gains in perceived agency and entertainment value without compromising coherence or safety boundaries. The work establishes personality as a tractable, measurable, and trainable dimension of AI behavior—opening pathways for more vivid, trustworthy, and human-resonant interactive agents.