Un-fine-tuned sparsified language models match baseline performance on reasoning tasks when evaluated using pass-at-4 or pass-at-8 sampling regimes