pair_preference_mapper.py

import re
from typing import Dict, Optional

from loguru import logger
from pydantic import PositiveInt

from data_juicer.ops.base_op import OPERATORS, Mapper
from data_juicer.utils.model_utils import get_model, prepare_model

OP_NAME = 'pair_preference_mapper'


# TODO: Extend LLM-based OPs into API-based implementation.
@OPERATORS.register_module(OP_NAME)
class PairPreferenceMapper(Mapper):
    """
    Mapper to construct paired preference samples.
    """

    # avoid leading whitespace
    DEFAULT_SYSTEM_PROMPT = (
        '你的任务是根据参考信息修改问答对中的回答，在语言风格、事实性、人物身份、立场等任一方面与原回答相反。'
        '必须按照以下标记格式输出，不要输出其他多余内容。\n'
        '【回答】\n'
        '生成的新回答\n'
        '【原因】\n'
        '生成该回答的原因')
    DEFAULT_INPUT_TEMPLATE = ('【参考信息】\n'
                              '{reference}\n'
                              '\n'
                              '以下是原始问答对：\n'
                              '【问题】\n'
                              '{query}\n'
                              '【回答】\n'
                              '{response}')
    DEFAULT_OUTPUT_PATTERN = r'.*?【回答】\s*(.*?)\s*【原因】\s*(.*)'

    def __init__(self,
                 api_model: str = 'gpt-4o',
                 *,
                 api_endpoint: Optional[str] = None,
                 response_path: Optional[str] = None,
                 system_prompt: Optional[str] = None,
                 input_template: Optional[str] = None,
                 output_pattern: Optional[str] = None,
                 rejected_key: str = 'rejected_response',
                 reason_key: str = 'reason',
                 try_num: PositiveInt = 3,
                 model_params: Dict = {},
                 sampling_params: Dict = {},
                 **kwargs):
        """
        Initialization method.

        :param api_model: API model name.
        :param api_endpoint: URL endpoint for the API.
        :param response_path: Path to extract content from the API response.
            Defaults to 'choices.0.message.content'.
        :param system_prompt: System prompt for guiding the generation task.
        :param input_template: Template for building the model input. It must
            contain placeholders '{query}' and '{reponse}', and can optionally
            include '{reference}'.
        :param output_pattern: Regular expression for parsing model output.
        :param rejected_key: The field name in the sample to store the
            generated rejected response. Defaults to 'rejected_response'.
        :param reason_key: The field name in the sample to store the reason for
            generating the response. Defaults to 'reason'.
        :param try_num: The number of retries for the API call in case of
            response parsing failure. Defaults to 3.
        :param model_params: Parameters for initializing the API model.
        :param sampling_params: Extra parameters passed to the API call.
            e.g {'temperature': 0.9, 'top_p': 0.95}
        :param kwargs: Extra keyword arguments.
        """
        super().__init__(**kwargs)

        self.system_prompt = system_prompt or self.DEFAULT_SYSTEM_PROMPT
        self.input_template = input_template or self.DEFAULT_INPUT_TEMPLATE
        self.output_pattern = output_pattern or self.DEFAULT_OUTPUT_PATTERN

        self.rejected_key = rejected_key
        self.reason_key = reason_key

        self.model_key = prepare_model(model_type='api',
                                       model=api_model,
                                       endpoint=api_endpoint,
                                       response_path=response_path,
                                       **model_params)
        self.try_num = try_num
        self.sampling_params = sampling_params

    def build_input(self, sample):
        mapping = {
            'query': sample[self.query_key],
            'response': sample[self.response_key],
            'reference': sample.get(self.text_key, '')
        }
        return self.input_template.format_map(mapping)

    def parse_output(self, raw_output):
        logger.debug(raw_output)
        match = re.match(self.output_pattern, raw_output, re.DOTALL)
        if match:
            return match.group(1).strip(), match.group(2).strip()
        else:
            return ('', '')

    def process_single(self, sample, rank=None):
        client = get_model(self.model_key, rank=rank)

        messages = [{
            'role': 'system',
            'content': self.system_prompt
        }, {
            'role': 'user',
            'content': self.build_input(sample)
        }]

        parsed_rejected, parsed_reason = '', ''
        for _ in range(self.try_num):
            try:
                output = client(messages, **self.sampling_params)
                parsed_rejected, parsed_reason = self.parse_output(output)
                if parsed_rejected and parsed_reason:
                    break
            except Exception as e:
                logger.warning(f'Exception: {e}')
        sample[self.rejected_key] = parsed_rejected
        sample[self.reason_key] = parsed_reason

        return sample