📘 概述

PerplexitySampleEvaluator 是一个基于 Hugging Face 语言模型计算文本困惑度（Perplexity）的评估算子。困惑度是衡量语言模型预测样本好坏程度的指标，其值越低，通常表示文本的流畅性和可理解性越高。该算子可用于评估生成文本的质量。

__init__函数

def __init__(self, model_name: str = 'gpt2', device='cuda'):

init参数说明

参数名	类型	默认值	说明
model_name	str	'gpt2'	Hugging Face Hub 上的模型名称或本地路径，用于计算困惑度。
device	str	'cuda'	模型运行的设备，例如 'cuda' 或 'cpu'。若不指定，则优先使用 CUDA。

Prompt模板说明

Prompt 模板名称	主要用途	适用场景	特点说明

run函数

def run(self, storage: DataFlowStorage, input_key: str = 'raw_content', output_key: str = 'PerplexityScore'):

参数

名称	类型	默认值	说明
storage	DataFlowStorage	必需	数据流存储实例，负责读取与写入数据。
input_key	str	"raw_content"	输入列名，对应待评估文本内容的字段。
output_key	str	"PerplexityScore"	输出列名，对应生成的困惑度分数字段。

🧠 示例用法

from dataflow.operators.text_pt.eval import PerplexitySampleEvaluator
from dataflow.utils.storage import FileStorage

# 准备数据和存储
storage = FileStorage(first_entry_file_name="pt_input.jsonl")

# 初始化并运行算子
perplexity_evaluator = PerplexitySampleEvaluator(model_name='gpt2')
perplexity_evaluator.run(
    storage.step(),
    input_key='raw_content',
    output_key='PerplexityScore'
)

🧾 默认输出格式（Output Format）

字段	类型	说明
raw_content	str	输入的待评估文本。
PerplexityScore	float	模型计算出的困惑度分数。

示例输入：

{
  "raw_content": "AMICUS ANTHOLOGIES, PART ONE (1965-1972)\nFebruary 23, 2017 Alfred Eaker Leave a comment\nWith Dr. Terror's House of Horrors (1965, directed by Freddie Francis and written by Milton Subotsky) Amicus Productions (spearheaded by Subotsky and Max Rosenberg, who previously produced for Hammer and was a cousin to Doris Wishman) established itself as a vital competitor to Hammer Studios..."
}

示例输出：

{
  "raw_content": "AMICUS ANTHOLOGIES, PART ONE (1965-1972)\nFebruary 23, 2017 Alfred Eaker Leave a comment\nWith Dr. Terror's House of Horrors (1965, directed by Freddie Francis and written by Milton Subotsky) Amicus Productions (spearheaded by Subotsky and Max Rosenberg, who previously produced for Hammer and was a cousin to Doris Wishman) established itself as a vital competitor to Hammer Studios...",
  "PerplexityScore": 49.2016410828
}