---
dataset_info:
  features:
  - name: query_id
    dtype: int64
  - name: query
    dtype: string
  - name: document
    dtype: string
  splits:
  - name: retail
    num_bytes: 16261464
    num_examples: 5000
  - name: videogames
    num_bytes: 7786542
    num_examples: 4360
  - name: books
    num_bytes: 2858945
    num_examples: 2245
  - name: news
    num_bytes: 11619385
    num_examples: 2375
  - name: web
    num_bytes: 17871918
    num_examples: 1500
  - name: debate
    num_bytes: 10085407
    num_examples: 880
  download_size: 33921309
  dataset_size: 66483661
configs:
- config_name: default
  data_files:
  - split: retail
    path: data/retail-*
  - split: videogames
    path: data/videogames-*
  - split: books
    path: data/books-*
  - split: news
    path: data/news-*
  - split: web
    path: data/web-*
  - split: debate
    path: data/debate-*
language:
- en
license: apache-2.0
tags:
- SEO
- CSEO
- RAG
- conversational-search-engine
---

## Dataset Summary

**C-SEO Bench** is a benchmark designed to evaluate conversational search engine optimization (C-SEO) techniques across two common tasks: **product recommendation** and **question answering**. Each task spans multiple domains to assess domain-specific effects and generalization ability of C-SEO methods.

## Supported Tasks and Domains

### Product Recommendation

This task requires an LLM to recommend the top-k products relevant to a user query, using only the content of 10 retrieved product descriptions. The task simulates a cold-start setting with no user profile. Domains:

- **Retail**: Queries and product descriptions from Amazon.
- **Video Games**: Search tags and game descriptions from Steam. 
- **Books**: GPT-generated queries with book synopsis from the Google Books API.

### Question Answering

This task involves answering queries based on multiple passages. Domains:

- **Web Questions**: Real search engine queries with retrieved web content.
- **News**: GPT-generated questions over sets of related news articles.
- **Debate**: Opinionated queries requiring multi-perspective evidence.

Total: Over **1.9k queries** and **16k documents** across six domains.

For more information about the dataset construction, please refer to the original publication.

Developed at [Parameter Lab](https://parameterlab.de/) with the support of [Naver AI Lab](https://clova.ai/en/ai-research).


## Disclaimer

> This repository contains experimental software results and is published for the sole purpose of giving additional background details on the respective publication.


## Citation
If this work is useful for you, please consider citing it

```
TODO
```

✉️ Contact person: Haritz Puerto, haritz.puerto@tu-darmstadt.de

🏢 https://www.parameterlab.de/

Don't hesitate to send us an e-mail or report an issue if something is broken (and it shouldn't be) or if you have further questions.