<p>High-Confidence Errors in Lightweight Instruction-Tuned Language Models</p>