Reframing Data Value for Large Language Models Through the Lens of Plausability