How to evaluate correctness and errors in multiple-choice tasks?

[WuXnkris]

"gen_args_0": {
"arg_0": "Question: a shopkeeper sold an article offering a discount of 5 % and earned a profit of 31.1 % . what would have been the percentage of profit earned if no discount had been offered ?\nAnswer:",
"arg_1": " 38"
},
"gen_args_1": {
"arg_0": "Question: a shopkeeper sold an article offering a discount of 5 % and earned a profit of 31.1 % . what would have been the percentage of profit earned if no discount had been offered ?\nAnswer:",
"arg_1": " 27.675"
},
"gen_args_2": {
"arg_0": "Question: a shopkeeper sold an article offering a discount of 5 % and earned a profit of 31.1 % . what would have been the percentage of profit earned if no discount had been offered ?\nAnswer:",
"arg_1": " 30"
}

............
"filtered_resps": [
["-7.680142164230347", "False"],
["-17.299633383750916", "False"],

Why are there two gen_args, and why are the resps numerical values?
How can we determine the correct answer generated by the LLM?
How to get the model’s original response?