如何正确定义基于GPT-3.5/LLama2的spacy-llm自定义无监督分类任务

核心需求明确 #

你需要从给定文本中提取目标实体集合，示例输入：

"The Basilica of San Petronio is a minor basilica and church of the Archdiocese of Bologna located in Bologna, Emilia Romagna, northern Italy."
期望输出：{"Basilica of San Petronio", "minor basilica"}

GPT-3.5 实现示例 #

1. 环境配置 #

确保安装依赖：

pip install spacy spacy-llm openai

2. spaCy 配置文件（config.cfg） #

[nlp]
lang = "en"
pipeline = ["llm"]

[components]

[components.llm]
factory = "llm"

[components.llm.model]
@llm_models = "spacy.GPT-3.5.v1"
api_key = "你的OpenAI API密钥"

[components.llm.task]
@llm_tasks = "spacy.TextCat.v3"
labels = []  # 无监督场景不预设标签
prompt = """从以下文本中提取所有属于教堂/宗教建筑类别的实体名称，返回格式为Python集合，只输出集合内容，不要其他解释：
文本：{{text}}"""

3. 运行代码 #

import spacy
from spacy_llm.util import assemble

def extract_entities(text):
    nlp = assemble("config.cfg")
    response = nlp.get_pipe("llm").model.predict([text])
    # 解析LLM返回的字符串为集合
    return eval(response[0])

result = extract_entities("The Basilica of San Petronio is a minor basilica and church of the Archdiocese of Bologna located in Bologna, Emilia Romagna, northern Italy.")
print(result)

Llama2 实现示例（本地部署） #

1. 环境配置 #

安装额外依赖：

pip install spacy spacy-llm transformers accelerate

2. spaCy 配置文件（config_llama2.cfg） #

[nlp]
lang = "en"
pipeline = ["llm"]

[components]

[components.llm]
factory = "llm"

[components.llm.model]
@llm_models = "spacy.Llama2.v1"
name = "meta-llama/Llama-2-7b-chat-hf"
token = "你的Hugging Face访问令牌"

[components.llm.task]
@llm_tasks = "spacy.TextCat.v3"
labels = []
prompt = """从以下文本中提取所有属于教堂/宗教建筑类别的实体名称，返回格式为Python集合，只输出集合内容，不要其他解释：
文本：{{text}}"""

3. 运行代码 #

import spacy
from spacy_llm.util import assemble

def extract_entities_llama(text):
    nlp = assemble("config_llama2.cfg")
    response = nlp.get_pipe("llm").model.predict([text])
    return eval(response[0])

result = extract_entities_llama("The Basilica of San Petronio is a minor basilica and church of the Archdiocese of Bologna located in Bologna, Emilia Romagna, northern Italy.")
print(result)

关键优化点 #

• Prompt 精准性：明确指定提取类别（教堂/宗教建筑）和输出格式（Python集合），避免LLM返回冗余内容
• 任务适配：无监督分类场景下，不要预设labels，让LLM完全根据Prompt提取目标内容
• 输出解析：直接解析LLM返回的字符串为Python集合，跳过TextCat默认的概率计算逻辑

优质参考方向 #

• spaCy-LLM官方文档：重点关注llm组件的自定义任务配置和模型适配章节
• Hugging Face Transformers与spaCy集成指南：学习本地LLM模型的加载和参数调优
• Prompt工程最佳实践：针对实体提取类任务，优化指令的明确性和约束性

内容的提问来源于stack exchange，提问作者ascimenti