CSConDa is the first Vietnamese QA dataset in the customer support domain. The dataset was curated and approved in collaboration with DooPage, a Vietnamese software company serving 30,000 customers and 45,000 advisors through its multi-channel support platform.

Covering a diverse range of topics, from product pricing inquiries to technical consultations, CSConDa serves as a representative dataset for this domain. The data is meticulously annotated following expert-guided annotation guidelines and is categorized into three types. The table below presents the categorization criteria.

No.	Question type	Definition	Requirements	Characteristics	Identification	Example from CSConDa
1	General	Common conversational phrases unrelated to a specific topic.	Typically greetings, farewells, or brief acknowledgments.	- Short with minimal semantic content.	Uses generic words, not tied to a specific issue or subject.	âu kê thank kiu e. (English: Okay, thank you.)
2	Simple	Direct, fact-based questions requiring concise responses.	Responses are short, providing direct or basic information.	- Concise and easy to understand. - Uses fewer linguistic variations than complex questions. - Less dependent on context.	Often includes service-related terms such as "Doopage", "chatbot", "Zalo", or other technical references.	Báo giá giúp M nhé. Gọi M lúc 11h số này nef <số điện thoại>. C cần qly 8 page, 2 Zalo, 1 website, 1 YTB, 3-5 người dùng. (English: Please send me the pricing details. Call me at 11 AM at this number <phone number>. I need 8 pages, 2 Zalo accounts, 1 website, and 1 YouTube channel for 3–5 users.)
3	Complex	Requires detailed explanations, often involving troubleshooting.	Responses must be elaborate, potentially including multiple steps or additional clarifications.	- Typically longer with more contextual details. - Often includes problem descriptions or error messages.	Frequently contains technical terms and detailed issue descriptions, often explaining why a problem occurs rather than just what it is.	à thế đây là teen maps. chứ đâu phải tên business đâu em. tên 1 địa chỉ map đó. Nhưng lsao mà gõ tên map vào đó được. trong khi bên trong cho phép add nhiều locations? Lỗi file này e ơi, ko down đc. Bên Zalo OA down bt. (English: Oh, so this is the name on Google Maps, not the business name. This is just a map location. But how can I enter a map name when multiple locations are allowed? This file is corrupted and cannot be downloaded. On Zalo OA, the download works fine.)

CSConDa can be used for fine-tuning Vietnamese LLMs, classification tasks, and benchmarking Vietnamese LLMs in the customer support domain.

This dataset is part of the work “A Benchmark Dataset and Evaluation Framework for Vietnamese Large Language Models in Customer Support”, accepted for publication in the Proceedings of the 17th International Conference on Computational Collective Intelligence (ICCCI 2025).

If you use CSConDa in your research, please cite:

@InProceedings{10.1007/978-3-032-10202-7_33,
author="Nguyen, Long S. T.
and Hua, Truong P.
and Nguyen, Thanh M.
and Pham, Toan Q.
and Ngo, Nam K.
and Nguyen, An X.
and Pham, Nghi D. M.
and Nguyen, Nghia H.
and Quan, Tho T.",
editor="Nguyen, Ngoc Thanh
and Dinh Duc Anh, Vu
and Kozierkiewicz, Adrianna
and Nguyen Van, Sinh
and Nunez, Manuel
and Treur, Jan
and Vossen, Gottfried",
title="A Benchmark Dataset and Evaluation Framework for Vietnamese Large Language Models in Customer Support",
booktitle="Advances in Computational Collective Intelligence",
year="2026",
publisher="Springer Nature Switzerland",
address="Cham",
pages="481--496",
abstract="With the rapid advancement of Artificial Intelligence, Large Language Models (LLMs) have become indispensable in Question Answering (QA) systems, enhancing response efficiency and reducing human workload, particularly in customer service. The rise of Vietnamese LLMs (ViLLMs) has positioned lightweight open-source models as the preferred choice due to their efficiency, accuracy, and privacy advantages. However, systematic evaluations of their performance in domain-specific contexts remain scarce, making it challenging for enterprises to identify the most suitable LLM for customer support applications, especially given the lack of benchmark datasets reflecting real-world customer interactions. To bridge this gap, we introduce Customer Support Conversations Dataset (CSConDa), a high-quality benchmark comprising over 9,000 QA pairs, meticulously curated from customer interactions with human advisors at a large-scale Vietnamese software company. Covering diverse service-related topics, including pricing inquiries, product availability, and technical troubleshooting, CSConDa serves as a representative dataset for evaluating ViLLMs in real-world scenarios. Furthermore, we present a comprehensive evaluation framework, benchmarking 11 lightweight open-source ViLLMs on CSConDa using not only well-suited automatic metrics but also an in-depth syntactic analysis to uncover their strengths, weaknesses, and underlying linguistic patterns. This analysis provides insights into model behavior, explains performance variations, and identifies critical areas for improvement, guiding future advancements in ViLLM development. Thus, by establishing a robust benchmark for LLM-driven customer service applications, our work provides a quantitative evaluation dataset and a comprehensive ViLLM performance comparison, offering key insights into intrinsic model performance, including accuracy, fluency, and consistency, while enabling informed decision-making for next-generation QA systems. Our dataset is publicly available on Hugging Face.",
isbn="978-3-032-10202-7"
}

Downloads last month: 29