Currently, users need to either define their own extract_label function or use the default one, which searches for the regular expression r"final answer:\s+(.*)".
However, for this to work properly and avoid random formatting from LLM outputs, we default to checking the lower case outputs when using the default extract_label. This can lead to failure to compute the accuracy if the groundtruth labels provided in datasets have capitalized letters. It is also suboptimal to ask every user to use lower case labels only.
There can be a quick fix, by changing the get_results function in hypogenic.utils.
Currently, users need to either define their own
extract_labelfunction or use the default one, which searches for the regular expressionr"final answer:\s+(.*)".However, for this to work properly and avoid random formatting from LLM outputs, we default to checking the lower case outputs when using the default
extract_label. This can lead to failure to compute the accuracy if the groundtruth labels provided in datasets have capitalized letters. It is also suboptimal to ask every user to use lower case labels only.There can be a quick fix, by changing the
get_resultsfunction inhypogenic.utils.