Bug: incorrect metrics evaluating for step two

[s-isaev]

https://github.com/microsoft/DeepSpeedExamples/blob/ab4e2e54620d0e80ead128b30dd39d9d55751eab/applications/DeepSpeed-Chat/training/step2_reward_model_finetuning/main.py#L263

Score and accuracy division should be out of for loop.

[s-isaev]

Here I proposed a fix: #426

[s-isaev]

https://github.com/microsoft/DeepSpeedExamples/blob/master/applications/DeepSpeed-Chat/training/step2_reward_model_finetuning/training_log_output/opt-350m_globalBatchSize-64.log

This log should be incorrect for now because of this bug.

[nepetune233]

@s-isaev Yeah, I also found the calculation is buggy #334 #347. But the accuracy should not be an issue, because it is calculated by correct_predictions / total_predictions, which will update in the loop.