Interruptive Prompting as Diagnostic Stress Testing of Local Large Language Models

This dataset accompanies the paper Interruptive Prompting as Diagnostic Stress Testing of Local Large Language Models and contains both the finalized manuscript and the raw interaction log used for analysis. The work documents a live, interrupt-driven Ollama session involving multiple locally hosted large language models (LLaMA-2 7B Chat, Phi-3 Mini, and Gemma 7B). By deliberately interrupting generation cycles and rapidly shifting prompt domains, the session functions as a lightweight behavioral stress test that reveals model-specific priors, alignment reflexes, verbosity biases, institutional mimicry, and refusal boundary geometry. Rather than evaluating benchmark accuracy or task performance, the study focuses on behavior under interruption, treating instability as signal rather than noise. The included interaction log serves as an executable empirical artifact supporting qualitative analysis in the paper. This dataset is intended for researchers interested in AI alignment diagnostics, local LLM evaluation, interpretability via interaction dynamics, and experimental methods that leverage adversarial play as a probe of model behavior. Files Interruptive Prompting As Diagnostic Stress Testing Of Local Large Language Models.pdf(58.67 KB, MD5: 181702f2f6c434efa7c93fa9433e1351)— Final academic manuscript. llm-local-log.md(MD5: 89a33b00cd7dfc271dd662a5063ec00f)— Raw Ollama session log used as the primary observational artifact.