---
license: apache-2.0
language:
- zh
tags:
- medical
- ecom
- video
- legal
size_categories:
- 10K<n<100K
configs:
- config_name: ecom
  data_files:
  - split: train
    path: qrelData/ecom_dev.tsv
- config_name: medical
  data_files:
  - split: train
    path: qrelData/medical_dev.tsv
- config_name: video
  data_files:
  - split: train
    path: qrelData/video_dev.tsv
- config_name: law
  data_files:
  - split: train
    path: qrelData/law_dev.tsv
---

# Dataset Card for Agentic Retrieval Benchmark

In recent years, there has been a large body of work in the field of text retrieval. However, existing studies are often scattered across different datasets, and their comparisons are partial and fragmented, lacking a comprehensive benchmark for evaluating retrieval performance.

This dataset is provided as part of the Agentic Retrieval Benchmark.

The project aims to establish a reproducible benchmark for LLM-augmented text retrieval. Under a unified dataset and pipeline, it compares the performance of traditional BM25 and vector-based retrieval methods, and evaluates the contribution of each component in the “query rewriting + vectorization” process.

## Dataset Description

The dataset is constructed from two open-source datasets:

* **Multi-CPR**: Covers three application scenarios (medical, e-commerce, and video), formatted as single-turn question-answer pairs.
* **LexRAG**: Focuses on Chinese legal consultation scenarios, formatted as multi-turn dialogues.

From the Multi-CPR dataset, we sample 1,000 queries and approximately 10,000 corpus passages for each of the three scenarios, using their corresponding indices as ground truth.

For the LexRAG dataset, we adopt a dialogue-history-plus-current-question setup. That is, each query consists of the full conversation history up to the current turn, combined with the latest user question.

The data has been cleaned and preprocessed, and can be directly used as input for query rewriting and evaluation scripts.

* Query data is stored in `.\data\rawData\xxx_query.txt`
* Passage data is stored in `.\data\rawData\xxx_subset.tsv`
* Ground truth labels/indices are stored in `.\data\qrelData\xxx_dev.tsv`

# Dataset Card for Agentic Retrieval Benchmark

近年来，在文本检索领域已有大量相关工作，但现有工作大多分散在不同的数据集上，且结果对比片面，缺少一个较为全面的文本检索性能指标测试基准。

本数据集为 Agentic Retrieval Benchmark 的配套数据。

项目旨在为大模型增强的文本检索建立可复现的基准：在统一的数据与流水线下比较传统 BM25 与向量检索的性能指标，并检验“重写 + 向量化”过程中各环节的贡献。

## Dataset Description

项目数据集来自 2 个开放数据集：
- Multi-CPR：包含三个应用场景（医疗、电商、视频），数据格式为单轮问答
- LexRAG：包含中文法律咨询场景，数据格式为多轮对话

我们从 Multi-CPR 数据集的 3 个场景中分别提取了 1000 条 query 和约 10000 条 corpus，并以对应的索引作为 groundtruth。

对于 LexRAG 数据集，我们选取了其中的对话历史+最新问题场景，即每次用于 query 的文本等于该次会话的前面所有轮次问答历史+当前最新一轮的提问。

数据已经完成清洗和预处理，可直接用作重写和评估脚本的输入。

- query 数据存放在`.\data\rawData\xxx_query.txt`

- passage 数据`.\data\rawData\xxx_subset.tsv`

- groundtruth 标签/索引存放在`.\data\qrelData\xxx_dev.tsv`