def regex_parser()

in src/alpaca_eval/completion_parsers.py [0:0]

• 16 lines of code
• 7 McCabe index (conditional complexity)

def regex_parser(completion: str, outputs_to_match: dict[Any, Any]) -> list[Any]:
    r"""Parse a single batch of completions, by returning a sequence of keys in the order in which outputs_to_match
    was matched.

    Parameters
    ----------
    completion : str
        Completion to parse.

    outputs_to_match : dict[str, Any]
        Dictionary of compiled regex to match. Keys are the keys to return in the order in which they are matched.
        The values can be either a compiled regex or a string. If a string, it will be compiled to a regex and that will
        be modified inplace.

    Examples
    --------
    >>> completion = ('\n(b)\n\n### Best output for example 8:\n(a)\n\n### Best output for example 9:\n(b)\n\n### Best'\
    ...               ' output for example 10:\n(a)\n\n### Best output for example 11:\n(a)')
    >>> regex_parser(completion, {1: r"\n\(a\)", 2: r"\n\(b\)"})
    [2, 1, 2, 1, 1]
    >>> regex_parser(' (a)', {1: r" \(a\)", 2: r" \(b\)"})
    [1]
    >>> completion = ('### Preferred output in JSON format for example 4:\r\n{{\r\n"Concise explanation": "Both'\
    ... ' outputs are incorrect, but Output (a) is less confusing and more concise.",\r\n"Output (a) is better than'\
    ... ' Output (b)": true\r\n}}\r\n\r\n### Preferred output in JSON format for example 5:\r\n{{\r\n"Concise'\
    ... ' explanation": "Both outputs are incomplete, but Output (b) seems to start with a more relevant source."'\
    ... ',\r\n"Output (a) is better than Output (b)": false\r\n}}\r\n\r\n### Preferred output in JSON format for'\
    ... ' example 6:\r\n{{\r\n"Concise explanation": "Both outputs are incorrect, but Output (a) is less confusing and'\
    ... ' more concise.",\r\n"Output (a) is better than Output (b)": true\r\n}}\r\n\r\n### Preferred output in JSON' \
    ... ' format for example 7:\r\n{{\r\n"Concise explanation": "Both outputs are incomplete, but Output (b) seems to'\
    ... ' start with a more relevant source.", \r\n"Output (a) is better than Output (b)": false\r\n}}')
    >>> regex_parser(completion, {1: ' true', 2: ' false'})
    [1, 2, 1, 2]
    """
    if completion == "":
        raise ValueError("The completion is empty.")

    for k, v in outputs_to_match.items():
        if not isinstance(v, re.Pattern):
            # inplace modification, which is bad practice but useful to speedup
            outputs_to_match[k] = re.compile(v)

    completion = copy.deepcopy(completion)
    responses = []
    while True:
        match, key = utils._find_first_match(completion, outputs_to_match)
        if not match:
            break
        responses.append(key)
        # avoid matching the same output twice
        completion = completion[match.end() :]
    return responses