Intent Ambiguity and Semantic Sanitization in a Local LLM: A "TRAIN" Double-Meaning Stress Test (Full Interaction Log)

This dataset contains a complete Markdown transcript documenting a conversational interaction between a locally hosted LLaMA-3.1-8B language model and a user employing sustained, playful ambiguity around the verb “train.” The user input intentionally maintains lexical consistency while escalating affect and cadence, introducing double-meaning without explicit sexual content. The interaction reveals a distinct model behavior pattern: semantic bleaching, affect absorption without intent acknowledgment, and activity-based reframing (e.g., trivia, word games) as an avoidance strategy. Rather than refusing or requesting clarification, the model suppresses ambiguity by enforcing a benign interpretive frame. The transcript is published without interpretive edits. The log itself constitutes the primary empirical artifact and is intended for use in studies of intent inference, conversational ambiguity, politeness bias, and human–AI interaction under suggestive but non-explicit language conditions.