diff --git a/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner.py b/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner.py index 06521176..0307bafd 100644 --- a/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner.py +++ b/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner.py @@ -11,6 +11,76 @@ from typing import Union import re + + +_CLEANED_BLOCK_RE = re.compile( + r"\s*(.*?)\s*", + flags=re.IGNORECASE | re.DOTALL, +) +_ANSWER_BLOCK_RE = re.compile( + r"\s*(.*?)\s*", + flags=re.IGNORECASE | re.DOTALL, +) +_PROCESSING_STEPS_RE = re.compile( + r"\s*Processing Steps:\s*" + r"1\.\s*\[Tag Analysis\]\s*Classify markup tags\s*" + r"2\.\s*\[Reference Extraction\]\s*Isolate images/tables\s*" + r"3\.\s*\[Character Audit\]\s*Log special chars\s*" + r"4\.\s*\[Structure Check\]\s*Validate hierarchy\s*" + r"5\.\s*\[Final Output\]\s*Generate cleaned text.*$", + flags=re.IGNORECASE | re.DOTALL, +) +_PROCESSING_STEPS_WITHOUT_HEADER_RE = re.compile( + r"\s*1\.\s*\[Tag Analysis\]\s*Classify markup tags\s*" + r"2\.\s*\[Reference Extraction\]\s*Isolate images/tables\s*" + r"3\.\s*\[Character Audit\]\s*Log special chars\s*" + r"4\.\s*\[Structure Check\]\s*Validate hierarchy\s*" + r"5\.\s*\[Final Output\]\s*Generate cleaned text.*$", + flags=re.IGNORECASE | re.DOTALL, +) +_ZH_PROCESSING_STEPS_RE = re.compile( + r"\s*(?:处理步骤|处理流程|治理步骤)[:：]\s*" + r"1[.、]\s*\[(?:标签分析|标记分析)\].*?" + r"2[.、]\s*\[(?:引用提取|参考提取)\].*?" + r"3[.、]\s*\[(?:字符审核|字符审计)\].*?" + r"4[.、]\s*\[(?:结构检查|结构校验)\].*?" + r"5[.、]\s*\[(?:最终输出|最终结果)\].*$", + flags=re.DOTALL, +) +_ZH_PROCESSING_STEPS_WITHOUT_HEADER_RE = re.compile( + r"\s*1[.、]\s*\[(?:标签分析|标记分析)\].*?" + r"2[.、]\s*\[(?:引用提取|参考提取)\].*?" + r"3[.、]\s*\[(?:字符审核|字符审计)\].*?" + r"4[.、]\s*\[(?:结构检查|结构校验)\].*?" + r"5[.、]\s*\[(?:最终输出|最终结果)\].*$", + flags=re.DOTALL, +) + + +def extract_cleaned_text(text, post_process=None) -> str: + """Extract model output while dropping prompt-instruction leakage.""" + text = "" if text is None else str(text) + + cleaned_match = _CLEANED_BLOCK_RE.search(text) + if cleaned_match: + text = cleaned_match.group(1) + else: + answer_match = _ANSWER_BLOCK_RE.search(text) + if answer_match: + text = answer_match.group(1) + text = text.replace("", "").replace("", "") + + text = _PROCESSING_STEPS_RE.sub("", text) + text = _PROCESSING_STEPS_WITHOUT_HEADER_RE.sub("", text) + text = _ZH_PROCESSING_STEPS_RE.sub("", text) + text = _ZH_PROCESSING_STEPS_WITHOUT_HEADER_RE.sub("", text) + text = text.strip() + + if post_process: + text = post_process(text) + return text + + @prompt_restrict( KnowledgeCleanerPrompt ) @@ -27,7 +97,7 @@ def __init__(self, llm_serving: LLMServingABC, lang="en", prompt_template : Unio if prompt_template: self.prompt_template = prompt_template else: - self.prompt_template = KnowledgeCleanerPrompt() + self.prompt_template = KnowledgeCleanerPrompt(lang=lang) @staticmethod def get_desc(lang: str = "zh"): @@ -123,15 +193,16 @@ def run( formatted_prompts = self._reformat_prompt(dataframe) cleaned = self.llm_serving.generate_from_input(formatted_prompts,"") - #for each in cleaned, only save the content in and + # Save only the final cleaned text, even if the model leaks prompt steps. cleaned_extracted = [ - str(text).split('')[1].split('')[0].strip() - if '' in str(text) and '' in str(text) - else str(text).strip() + extract_cleaned_text( + text, + getattr(self.prompt_template, "_post_process", None), + ) for text in cleaned ] dataframe[self.output_key] = cleaned_extracted output_file = storage.write(dataframe) self.logger.info(f"Results saved to {output_file}") - return [output_key] \ No newline at end of file + return [output_key] diff --git a/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner_batch.py b/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner_batch.py index 53a445f6..cd386319 100644 --- a/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner_batch.py +++ b/dataflow/operators/knowledge_cleaning/generate/kbc_text_cleaner_batch.py @@ -1,4 +1,5 @@ from dataflow.prompts.kbcleaning import KnowledgeCleanerPrompt +from dataflow.operators.knowledge_cleaning.generate.kbc_text_cleaner import extract_cleaned_text import pandas as pd from dataflow.utils.registry import OPERATOR_REGISTRY from dataflow import get_logger @@ -125,11 +126,12 @@ def run( raw_chunks, formatted_prompts = self._reformat_prompt_from_path(chunk_path) cleaned = self.llm_serving.generate_from_input(formatted_prompts, "") - # for each in cleaned, only save the content in and + # Save only the final cleaned text, even if the model leaks prompt steps. cleaned_extracted = [ - text.split('')[1].split('')[0].strip() - if '' in str(text) and '' in str(text) - else str(text).strip() + extract_cleaned_text( + text, + getattr(self.prompts, "_post_process", None), + ) for text in cleaned ] json_items=[{ diff --git a/dataflow/prompts/kbcleaning.py b/dataflow/prompts/kbcleaning.py index 16067fb8..47010bba 100644 --- a/dataflow/prompts/kbcleaning.py +++ b/dataflow/prompts/kbcleaning.py @@ -177,35 +177,17 @@ def build_prompt(self, raw_content: str) -> str: """ if self.lang == "en": - processing_steps = """ -Processing Steps: -1. [Tag Analysis] Classify markup tags -2. [Reference Extraction] Isolate images/tables -3. [Character Audit] Log special chars -4. [Structure Check] Validate hierarchy -5. [Final Output] Generate cleaned text -""".strip() - output_requirement = 'Response must contain ONLY cleaned text between and .' + output_requirement = 'Response must contain ONLY the final cleaned text between and . Do not include analysis, explanations, processing steps, or bullet lists outside the cleaned text.' else: - processing_steps = """ -处理步骤： -1. [标签分析] 识别并分类所有标记标签 -2. [引用提取] 分离图片/表格/签名等引用内容 -3. [字符审核] 记录特殊字符变更 -4. [结构检查] 验证文本层级 -5. [最终输出] 生成清洗后文本 -""".strip() output_requirement = '响应必须只包含清洗后文本，以开头，结尾，无其他内容。' return f""" {self.prompt_header} +{output_requirement} + 待清洗内容： {raw_content} - -{processing_steps} - -{output_requirement} """.strip() def _post_process(self, cleaned_text: str) -> str: