generate_qa_from_examples_mapper.py

import json
import random
import re
from typing import Dict, Optional

from loguru import logger
from pydantic import PositiveInt

from data_juicer.utils.lazy_loader import LazyLoader
from data_juicer.utils.model_utils import get_model, prepare_model

from ..base_op import OPERATORS, Mapper

torch = LazyLoader('torch', 'torch')
vllm = LazyLoader('vllm', 'vllm')
rouge = LazyLoader('rouge', 'rouge')

OP_NAME = 'generate_qa_from_examples_mapper'


# TODO: Extend LLM-based OPs into API-based implementation.
@OPERATORS.register_module(OP_NAME)
class GenerateQAFromExamplesMapper(Mapper):
    """
    Mapper to generate question and answer pairs from examples.
    You should configure an empty dataset in your yaml config file:
    ```
    generated_dataset_config:
      type: 'EmptyFormatter'  # use `RayEmptyFormatter` when enable ray
      length: ${The number of generated samples}
      feature_keys: ${text key}
    ```
    The number of samples generated is determined by
    the length of the empty dataset.
    """

    DEFAULT_SYSTEM_PROMPT = (
        '请你仔细观察多个示例数据的输入和输出，按照你的理解，总结出相应规矩，然后写出一个新的【问题】和【回答】。'
        '注意，新生成的【问题】和【回答】需要满足如下要求：\n'
        '1. 生成的【问题】和【回答】不能与输入的【问题】和【回答】一致，但是需要保持格式相同。\n'
        '2. 生成的【问题】不一定要局限于输入【问题】的话题或领域，生成的【回答】需要正确回答生成的【问题】。\n'
        '3. 提供的【问题】和【回答】可能是多轮对话，生成的【问题】和【回答】也可以是多轮，但是需要保持格式相同。\n'
        '4. 生成的【问题】和【回答】必须成对出现，而且【问题】需要在【回答】之前。\n')

    DEFAULT_INPUT_TEMPLATE = '{}'
    DEFAULT_EXAMPLE_TEMPLATE = '\n如下是一条示例数据：\n{}'
    DEFAULT_QA_PAIR_TEMPLATE = '【问题】\n{}\n【回答】\n{}\n'
    DEFAULT_OUTPUT_PATTERN = r'【问题】(.*?)【回答】(.*?)(?=【问题】|$)'

    _accelerator = 'cuda'

    def __init__(self,
                 hf_model: str = 'Qwen/Qwen2.5-7B-Instruct',
                 *,
                 seed_file: str = '',
                 example_num: PositiveInt = 3,
                 similarity_threshold: float = 0.7,
                 system_prompt: Optional[str] = None,
                 input_template: Optional[str] = None,
                 example_template: Optional[str] = None,
                 qa_pair_template: Optional[str] = None,
                 output_pattern: Optional[str] = None,
                 enable_vllm: bool = False,
                 model_params: Optional[Dict] = None,
                 sampling_params: Optional[Dict] = None,
                 **kwargs):
        """
        Initialization method.

        :param hf_model: Hugginface model ID.
        :param seed_file: Path to the seed file in chatml format.
        :param example_num: The number of selected examples.
            Randomly select N examples from "seed_file" and
            put them into prompt as QA examples.
        :param similarity_threshold: The similarity score threshold
            between the generated samples and the seed examples.
            Range from 0 to 1. Samples with similarity score less than
            this threshold will be kept.
        :param system_prompt: System prompt for guiding the generation task.
        :param input_template: Template for building the input prompt. It must
            include one placeholder '{}', which will be replaced by
            `example_num` formatted examples defined by `example_template`.
        :param example_template: Template for formatting one QA example. It
            must include one placeholder '{}', which will be replaced by one
            formatted qa_pair.
        :param qa_pair_template: Template for formatting a single QA pair
            within each example. Must include two placeholders '{}' for the
            question and answer.
        :param output_pattern: Regular expression pattern to extract questions
            and answers from model response.
        :param enable_vllm: Whether to use vllm for inference acceleration.
        :param model_params: Parameters for initializing the model.
        :param sampling_params: Sampling parameters for text generation.
            e.g {'temperature': 0.9, 'top_p': 0.95}
        :param kwargs: Extra keyword arguments.
        """
        super().__init__(**kwargs)

        if not seed_file:
            raise ValueError(
                'Please provide `seed_file` in chatml format.'
                'Example: data-juicer/demos/data/demo-dataset-chatml.jsonl')

        self.seed_file = seed_file
        self.example_num = example_num
        self.similarity_threshold = similarity_threshold
        self.similarity_type = 'rouge_l'

        self.system_prompt = system_prompt or self.DEFAULT_SYSTEM_PROMPT
        self.input_template = input_template or self.DEFAULT_INPUT_TEMPLATE
        self.example_template = example_template or self.DEFAULT_EXAMPLE_TEMPLATE  # noqa: E501
        self.qa_pair_template = qa_pair_template or \
            self.DEFAULT_QA_PAIR_TEMPLATE
        self.output_pattern = output_pattern or self.DEFAULT_OUTPUT_PATTERN

        self.enable_vllm = enable_vllm
        model_params = model_params or {}
        sampling_params = sampling_params or {}

        if enable_vllm:
            assert torch.cuda.device_count() >= 1, 'must be executed in CUDA'
            # cannot initialize vllm replicas on different GPUs
            self.num_proc = 1
            if model_params.get('tensor_parallel_size') is None:
                tensor_parallel_size = torch.cuda.device_count()
                logger.info(f'Set tensor_parallel_size to \
                    {tensor_parallel_size} for vllm.')
                model_params['tensor_parallel_size'] = tensor_parallel_size
            self.model_key = prepare_model(
                model_type='vllm',
                pretrained_model_name_or_path=hf_model,
                **model_params)
            self.sampling_params = vllm.SamplingParams(**sampling_params)
        else:
            self.model_key = prepare_model(
                model_type='huggingface',
                pretrained_model_name_or_path=hf_model,
                return_pipe=True,
                **model_params)
            self.sampling_params = sampling_params

        self.seed_qa_samples = self._load_seed_qa_samples()
        if len(self.seed_qa_samples) == 0:
            raise ValueError('No QA data was parsed from the seed file!')

    def _load_seed_qa_samples(self):
        """Load QA pairs from chatml format file."""
        qa_samples = []
        with open(self.seed_file, encoding='utf-8') as f:
            lines = f.readlines()
            for line in lines:
                line = line.strip()
                qa_pairs = self._parse_chatml_str(line)
                if len(qa_pairs) > 0:
                    qa_samples.append(qa_pairs)
        return qa_samples

    def _sample_to_str(self, qa_sample):
        return '\n'.join(['\n'.join(qa_pair) for qa_pair in qa_sample]) + '\n'

    def _max_rouge_l_score(self, hypothesis, references):
        r = rouge.Rouge()
        max_score = 0.0
        hyp_str = self._sample_to_str(hypothesis)
        for reference in references:
            ref_str = self._sample_to_str(reference)
            scores = r.get_scores(hyp_str, ref_str)
            rouge_l_score = scores[0]['rouge-l']['f']
            if rouge_l_score > max_score:
                max_score = rouge_l_score
        return max_score

    def _parse_chatml_str(self, sample_str):
        user_input = None
        assistant_output = None
        qa_pairs = []
        data = json.loads(sample_str)
        for message in data['messages']:
            role = message['role']
            content = message['content']
            if role == 'user':
                user_input = content
            elif role == 'assistant':
                assistant_output = content
                qa_pairs.append((user_input, assistant_output))
        return qa_pairs

    def build_input(self, qa_examples):

        def format_qa_pairs(qa_example):
            return ''.join([
                self.qa_pair_template.format(q, a) for q, a in qa_example
                if q and a
            ])

        formatted_examples = ''.join([
            self.example_template.format(qa_pairs=format_qa_pairs(qa_example))
            for qa_example in qa_examples
        ])
        input_prompt = self.input_template.format(examples=formatted_examples)
        return input_prompt

    def parse_output(self, raw_output):
        logger.debug(raw_output)
        output_qa_pairs = []
        matches = re.findall(self.output_pattern, raw_output, re.DOTALL)
        for match in matches:
            question, answer = match
            output_qa_pairs.append((question.strip(), answer.strip()))
        return output_qa_pairs

    def process_single(self, sample, rank=None):
        model, _ = get_model(self.model_key, rank, self.use_cuda())

        random_qa_samples = random.sample(self.seed_qa_samples,
                                          self.example_num)
        input_prompt = self.build_input(random_qa_samples)

        messages = [{
            'role': 'system',
            'content': self.system_prompt
        }, {
            'role': 'user',
            'content': input_prompt
        }]

        if self.enable_vllm:
            response = model.chat(messages, self.sampling_params)
            output = response[0].outputs[0].text
        else:
            # model is pipe
            response = model(messages,
                             return_full_text=False,
                             **self.sampling_params)
            output = response[0]['generated_text']

        output_qa_pairs = self.parse_output(output)
        if len(output_qa_pairs) == 0:
            logger.warning('Parse model response error! '
                           'No data generated for the current response!')
            sample.update({
                self.query_key: '',
                self.response_key: '',
                self.history_key: self.empty_history()
            })
            return sample

        if self.similarity_type == 'rouge_l':
            sim_score = self._max_rouge_l_score(output_qa_pairs,
                                                random_qa_samples)
        else:
            raise ValueError(
                f'Not support similarity type "{self.similarity_type}"!')

        if sim_score <= self.similarity_threshold:
            query, response = output_qa_pairs[-1]
            history = output_qa_pairs[:-1]
            if len(history) == 0:
                history = self.empty_history()
        else:
            query = response = ''
            history = self.empty_history()
            logger.info('Filter this generated sample due to similarity.')

        sample.update({
            self.query_key: query,
            self.response_key: response,
            self.history_key: history
        })
        return sample