Not able to produce result as mentioned in the paper

Hi,
I found result on STS Benchmark without training sbert -
sbert model include bert layer + pooling layer and I evaluate it using sentEval.
and the same evaluation on USE. Below are the results -

**USE**
{'STSBenchmark': {'devpearson': 0.8195724907058578, 'pearson': 0.7873882116295602, 'spearman': 0.7808647077439052, 'mse': 1.176961038436245, 'yhat': array([2.08647348, 1.91908006, 1.55792314, ..., 3.86090869, 3.46508917, 3.56250464]), 'ndev': 1500, 'ntest': 1379}}

**SBERT Without any training data**
{'STSBenchmark': {'devpearson': 0.7133810714675861, 'pearson': 0.6539654665152129, 'spearman': 0.6447501570280392, 'mse': 1.4892911081516527, 'yhat': array([1.36104402, 1.66169352, 2.40927465, ..., 3.96148573, 3.95960486, 3.64619382]), 'ndev': 1500, 'ntest': 1379}}

Start by reviewing the reported SentEval STSBenchmark evaluation and the SBERT configuration described in the issue, including the BERT and pooling layers without training. Compare that setup with the paper's evaluation procedure and determine whether the reported result can be reproduced; document the cause of any discrepancy.

Written by the indexing model from the issue text.