大模型生成RAG评估数据集并计算hit_rate 和 mrr

背景

最近在做RAG评估的实验，需要一个RAG问答对的评估数据集。在网上没有找到好用的，于是便打算自己构建一个数据集。

简介

本文使用大模型自动生成RAG 问答数据集。使用BM25关键词作为检索器，然后在问答数据集上评估该检索器的效果。
输入是一篇文本，使用llamaindex加载该文本，使用prompt让大模型针对输入的文本生成提问。
步骤如下：

1. llamaindex 加载数据；
2. 利用 chatglm3-6B 构建CustomLLM；
3. 使用prompt和chatglm，结合文本生成对应的问题，构建RAG问答数据集；
4. 使用BM25Retriever，构建基于关键词的检索器；
5. 评估BM25Retriever在数据集上的hite_rate和mrr结果；

由于在构建问答对时，让大模型结合文本生成对应的问题。笔者在测试时，发现关键词检索比向量检索效果要好

代码实现

导入包

from typing import List, Any

from llama_index.core import SimpleDirectoryReader

from llama_index.core.node_parser import SentenceWindowNodeParser
from llama_index.legacy.llms import (
    CustomLLM, CompletionResponse, CompletionResponseGen, LLMMetadata)
from llama_index.legacy.schema import NodeWithScore, QueryBundle, Node
from llama_index.core.base.base_retriever import BaseRetriever
from llama_index.legacy.retrievers import BM25Retriever
from llama_index.core.evaluation import RetrieverEvaluator
from llama_index.core.evaluation import (
    generate_question_context_pairs,
    EmbeddingQAFinetuneDataset,
)

加载数据，使用llamaindex网站的paul_graham_essay.txt

# Load data
documents = SimpleDirectoryReader(
    input_files=["data/paul_graham_essay.txt"]
).load_data()

# create the sentence window node parser w/ default settings
node_parser = SentenceWindowNodeParser.from_defaults(
    window_size=3,
    window_metadata_key="window",
    original_text_metadata_key="original_text",
)

# Extract nodes from documents
nodes = node_parser.get_nodes_from_documents(documents)

# by default, the node ids are set to random uuids. To ensure same id's per run, we manually set them.
for idx, node in enumerate(nodes):
    node.id_ = f"node_{idx}"

大模型加载
chatglm3-6B 使用half，显存占用12G

from modelscope import snapshot_download
from modelscope import AutoTokenizer, AutoModel

model_name = "chatglm3-6b"
model_path = snapshot_download('ZhipuAI/chatglm3-6b')
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()
model = model.eval()

本地自定义大模型

# set context window size
context_window = 2048
# set number of output tokens
num_output = 256


class ChatGML(CustomLLM):
    @property
    def metadata(self) -> LLMMetadata:
        """Get LLM metadata."""
        return LLMMetadata(
            context_window=context_window,
            num_output=num_output,
            model_name=model_name,
        )

    def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
        prompt_length = len(prompt)

        # only return newly generated tokens
        text, _ = model.chat(tokenizer, prompt, history=[])
        return CompletionResponse(text=text)

    def stream_complete(
            self, prompt: str, **kwargs: Any
    ) -> CompletionResponseGen:
        raise NotImplementedError()


llm_model = ChatGML()

生成RAG测试数据集

# Prompt to generate questions
qa_generate_prompt_tmpl = """\
Context information is below.

---------------------
{context_str}
---------------------

Given the context information and not prior knowledge.
generate only questions based on the below query.

You are a Professor. Your task is to setup \
{num_questions_per_chunk} questions for an upcoming \
quiz/examination. The questions should be diverse in nature \
across the document. The questions should not contain options, not start with Q1/ Q2. \
Restrict the questions to the context information provided.\
"""
# The questions should be solely based on the provided context information, and please pose them in Chinese.\

qa_dataset = generate_question_context_pairs(
    nodes,
    llm=llm_model,
    num_questions_per_chunk=2,
    qa_generate_prompt_tmpl=qa_generate_prompt_tmpl
)

qa_dataset.save_json("pg_eval_dataset.json")
# qa_dataset = EmbeddingQAFinetuneDataset.from_json("pg_eval_dataset.json")

import pandas as pd

def display_results(eval_results):
    """
        计算hit_rate和mrr的平均值
    """

    metric_dicts = []
    for eval_result in eval_results:
        metric_dict = eval_result.metric_vals_dict
        metric_dicts.append(metric_dict)

    full_df = pd.DataFrame(metric_dicts)

    hit_rate = full_df["hit_rate"].mean()
    mrr = full_df["mrr"].mean()

    metric_df = pd.DataFrame(
        {"hit_rate": [hit_rate], "mrr": [mrr]}
    )
    return metric_df

class JieRetriever(BM25Retriever, BaseRetriever):
    def _get_scored_nodes(self, query: str):
        tokenized_query = self._tokenizer(query)
        doc_scores = self.bm25.get_scores(tokenized_query)
        nodes = []
        for i, node in enumerate(self._nodes):
            node_new = Node.from_dict(node.to_dict())
            node_with_score = NodeWithScore(node=node_new, score=doc_scores[i])
            nodes.append(node_with_score)
        return nodes
    
    def _retrieve(self, query_bundle: QueryBundle) -> List[NodeWithScore]:
        if query_bundle.custom_embedding_strs or query_bundle.embedding:
            logger.warning("BM25Retriever does not support embeddings, skipping...")

        scored_nodes = self._get_scored_nodes(query_bundle.query_str)

        # Sort and get top_k nodes, score range => 0..1, closer to 1 means more relevant
        nodes = sorted(scored_nodes, key=lambda x: x.score or 0.0, reverse=True)
        return nodes[:self._similarity_top_k]

retriever = JieRetriever.from_defaults(
# retriever = BM25Retrieve r.from_defaults(
                            nodes=nodes,
                            similarity_top_k=10
                        )

现在llamaindex在使用BM25Retrieve会报错，故笔者创建了JieRetriever，具体请点击查看链接

from llama_index.core.base.base_retriever import BaseRetriever

retriever_evaluator = RetrieverEvaluator.from_metric_names(
            ["mrr", "hit_rate"], retriever=retriever
        )
eval_results = await retriever_evaluator.aevaluate_dataset(qa_dataset)

for idx, item in enumerate(eval_results):
    if idx == 15:
        break
    d = item.metric_vals_dict
    mrr, hit_rate = d['mrr'], d['hit_rate']
    if mrr != 1 or hit_rate != 1:
        print(mrr, hit_rate, item.expected_ids, item.retrieved_ids)

下图展示了hit_rate 和 mrr 的计算：

结合下述结果，分析一下 hit_rate 和 mrr：

0.5 1.0 ['node_2'] ['node_71', 'node_2', 'node_0', 'node_199', 'node_126', 'node_419', 'node_446', 'node_218', 'node_1', 'node_70']

• ['node_2'] 是 label
• ['node_71', 'node_2', 'node_0', 'node_199', 'node_126', 'node_419', 'node_446', 'node_218', 'node_1', 'node_70'] 是检索器召回的候选列表；
• mrr : 0.5；'node_2' 在候选列表的第二个位置，故mrr为 二分之一。在第几位就是几分之一；
• hit_rate：代表label是否在候选集中，在就是1，不在就是0；

def display_results(eval_results):
    """
    	计算平均 hit_rate 和 mrr
    """

    metric_dicts = []
    for eval_result in eval_results:
        metric_dict = eval_result.metric_vals_dict
        metric_dicts.append(metric_dict)

    full_df = pd.DataFrame(metric_dicts)

    hit_rate = full_df["hit_rate"].mean()
    mrr = full_df["mrr"].mean()

    metric_df = pd.DataFrame(
        {"hit_rate": [hit_rate], "mrr": [mrr]}
    )
    return metric_df

display_results(eval_results)

公开

生成的评估数据集和相应示例代码，已上传到modelscope平台；

https://www.modelscope.cn/datasets/jieshenai/paul_graham_essay_rag/files

下一步阅读

在同样的数据集上，下一篇文章，展示了如何使用向量检索和rerank模型；

• 使用向量检索和rerank 在RAG数据集上实验评估hit_rate和mrr
  

参考资料

• https://www.llamaindex.ai/blog/boosting-rag-picking-the-best-embedding-reranker-models-42d079022e83

• LLM之RAG实战（二十五）| 使用LlamaIndex和BM25重排序实践

• TorchV的RAG实践分享(二)：基于ElasticSearch的混合检索实战&原理分析