Datasets:

Mer1Alii
/

TR-ECommerce-CustomerSupport-Instructions

license: apache-2.0
task_categories:
  - text-generation
language:
  - tr
tags:
  - e-commerce
  - customer-support
  - chain-of-thought
  - instruction-tuning
pretty_name: TR E-Commerce Customer Support Instructions
size_categories:
  - n<1K

TR E-Commerce Customer Support Instructions 🇹🇷

A high-quality Turkish E-Commerce Customer Support dataset designed for fine-tuning large language models (LLMs) on instruction-following customer support tasks.

Dataset Summary

Feature	Value
Language	Turkish (tr)
Domain	E-Commerce Customer Support
Format	Conversation (Conversational/Chat)
Chain-of-Thought (CoT)	✅ Natural paragraph reasoning (`thinking` field)
Total Categories	20
Total Rows	186 conversations

Categories & Distribution

#	Category	Description	Count
1	Sipariş Oluşturma	Cart errors, page load issues during checkout, guest checkout	10
2	Sipariş İptali	Order cancellation before/after shipping, partial cancellation	10
3	Sipariş Takibi	Status tracking, order processing delays, missing confirmation	10
4	Kargo Gecikmesi	Delivery delay, package stuck in transit, express delivery issues	8
5	Teslimat Sorunlari	Package marked delivered but not received, left with neighbor	10
6	Eksik Ürün	Missing items in order, missing accessories inside the box	8
7	Yanlis Ürün	Wrong color, size, or completely incorrect model delivered	8
8	Hasarli Ürün	Broken/cracked items, damaged packaging, used product received	8
9	İade	Return request, right of withdrawal, refund status, return shipping	12
10	Değişim	Size exchange, color exchange, product exchange process	10
11	Ödeme	3D Secure failures, double charges, installments, bank transfer	12
12	Kupon	Coupon code errors, expired codes, minimum cart value rules	8
13	Kampanya	Discounts not applied, campaign rules, first-order discounts	10
14	Hesap	Login issues, account deletion (GDPR/KVKK), account hijacking	10
15	Şifre	Password reset steps, verification email not arriving	8
16	Adres Değişikliği	Changing address after shipment, updating registration address	8
17	Fatura	Invoice details correction, corporate invoice, PDF download	8
18	Satıcı İletişimi	Seller not responding, counterfeit suspicion, seller disputes	10
19	Ürün Değerlendirmesi	Product authenticity, stock query, size charts, descriptions	10
20	Garanti	Warranty period, warranty coverage, authorized service centers	8

Data Format

Each conversation is formatted as a list of message dictionaries featuring role, content, images, thinking (for the assistant), and tool_calls fields:

[
  {
    "role": "user",
    "content": "merhaba siparişim nerede kaldı acaba",
    "images": null,
    "thinking": null,
    "tool_calls": null
  },
  {
    "role": "assistant",
    "content": "Merhaba, hemen kontrol edelim. Siparişinizin durumunu sorgulayabilmem için e-posta adresinizi veya takip numaranızı paylaşabilir misiniz?",
    "images": null,
    "thinking": "Kullanıcı siparişinin nerede olduğunu merak ediyor. Muhtemelen kargoya verilmiş ama takip bilgisi güncellenmemiş olabilir. Kargo takip numarasıyla kontrol etmesini önermeli ya da sipariş bilgisini alıp ben kontrol etmeliyim.",
    "tool_calls": null
  }
]

Dataset Key Design Features

Natural User Messages: Users write in realistic, natural Turkish. Some messages contain missing punctuation or all-lowercase letters (mimicking actual human messaging habits), without using excessive chat slangs or abbreviations.
Diverse Assistant Openings: Assistant responses use more than 10 different opening variations to avoid repetitive patterns in LLM fine-tuning.
Fluid Chain-of-Thought (CoT): Assistant entries include a single-paragraph thinking block containing the logical reasoning flow before producing the final response.
Varied Information Requests: Unlike rigid datasets asking only for order numbers, the assistant requests different details depending on the scenario (e.g., e-mail, phone number, tracking ID, screenshot) or resolves the query directly without asking for any information.
Length Constraints: Assistant responses are kept concise (ranging from 100 to 170 words) to mimic realistic customer support agent behaviors.

Usage

You can easily load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("Mer1Alii/TR-ECommerce-CustomerSupport-Instructions")
print(dataset["train"][0])

Fine-Tuning Tools

This dataset is compatible with popular training frameworks such as:

License

This dataset is licensed under the Apache 2.0 License. It is intended for educational, research, and commercial applications.

Developer

Mert Ali Alkan