Fixes #14.

check_currentmetric builds the inputs for the accuracy and f1 metrics as

predictions = [0 if prediction == reference else 1 for prediction, reference in zip(predictions, references)]
references = [1] * len(references)

and passes them to metrics/accuracy and metrics/f1. accuracy_score(references, predictions) then counts the rows where the prediction is 1, which are the rows where the generation did not match. A correct generation scores 0 and a wrong or empty one scores 1.

This swaps the two values, so a match is 1. The change is the same one line in codefuseEval/execution.py and in codefuseEval_202503/execution.py. The f1 value keeps its existing formula (F1 of the positive class with all-1 references); only its polarity changes.

Verification

The repo has no Python tests, so I did not add one. I extracted check_currentmetric verbatim from each file with ast and ran it against the repo's own metrics/accuracy/accuracy.py and metrics/f1/f1.py with real scikit-learn 1.4.2. evaluate and datasets are small stand-ins (only the Metric base class and a docstring decorator). Each row is a set of 10 generations checked against the reference "x":

codefuseEval_202503 has no metrics/ folder in this repository, so I ran its check_currentmetric (byte-identical to the other copy) against the metrics/ folder from codefuseEval; the results are the same as above.