In their readme they mention `output: str, the answer to the instruction as generated by text-davinci-003`.

And in the alpaca-data file there's this:

    {
        "instruction": "Perform the following arithmetic operation.",
        "input": "(5+7)\*3.",
        "output": "60."
    },

Do we need human QA on the training data? It's contaminating the set! I wonder how incorrect data like this will be filtered out in the future. Unless we're ok with that kind of new math.