Datasets:
Download README.md from ShinRunner1st/ServiceHub_Dataset_14Cat: direct link, hf CLI and curl.
- Browser
- Download file 8.41 kB
-
https://huggingface.co/datasets/ShinRunner1st/ServiceHub_Dataset_14Cat/resolve/main/README.md
- Command line
-
hf download hf://datasets/ShinRunner1st/ServiceHub_Dataset_14Cat/README.md
-
curl -L -o README.md https://huggingface.co/datasets/ShinRunner1st/ServiceHub_Dataset_14Cat/resolve/main/README.md
language:
- th
license: cc-by-4.0
task_categories:
- text-classification
task_ids:
- multi-class-classification
size_categories:
- 1K<n<10K
tags:
- thai
- it-service-request
- helpdesk
- synthetic
- llm-generated
pretty_name: Servicehub Thai IT Service Requests (14 categories)
ServiceThub Thai IT Service Requests — 14 Categories
ชุดข้อมูลคำร้องขอบริการด้าน IT และงานสำนักงานภาษาไทย สำหรับงานจำแนกหมวดหมู่ข้อความ (text classification) ปรับปรุงจากชุดข้อมูลต้นฉบับ 18 หมวด โดยตัดหมวดที่นิยามไม่ชัดเจนออกและรวมหมวดที่ขอบเขตทับซ้อนกัน เหลือ 14 หมวด
A Thai-language dataset of IT / office service requests for multi-class text classification. Derived from an 18-category version by dropping the ill-defined Other class and merging three pairs of overlapping classes, resulting in 14 categories.
⚠️ ข้อความทั้งหมดถูกสร้างด้วย LLM (synthetic data) และ label มาจากหมวดที่กำหนดใน prompt ตอนสร้างข้อความ ไม่ได้ผ่านการแท็กโดยมนุษย์ โปรดอ่านหัวข้อ ข้อจำกัด ก่อนนำไปใช้
ภาพรวม (Dataset Summary)
| รายการ | ค่า |
|---|---|
| จำนวนแถว | 2,850 |
| จำนวนหมวด (category) | 14 |
| ภาษา | ไทย (มีคำภาษาอังกฤษปนตามบริบทงาน IT) |
| ความยาวข้อความ | เฉลี่ย ~109 ตัวอักษร (ต่ำสุด 8, สูงสุด 283) |
| แถวซ้ำ | ไม่มี |
| Split | train เพียง split เดียว (ผู้ใช้แบ่ง train/test เอง) |
โครงสร้างข้อมูล (Data Fields)
| คอลัมน์ | ชนิด | คำอธิบาย |
|---|---|---|
request_text |
string | ข้อความคำร้องขอบริการภาษาไทย |
category |
string | หมวดหมู่ของคำร้อง (14 หมวด) — label หลัก |
priority |
string | ระดับความเร่งด่วน: High, Medium, Low |
sentiment |
string | อารมณ์ของข้อความ: Positive, Neutral, Negative |
_generation |
string | metadata ตอนสร้างข้อมูล (dict ในรูป string) มี target_category, target_priority, target_sentiment |
หมายเหตุ: target_category ใน _generation เป็นชื่อหมวดเดิมแบบ 18 หมวด ก่อนการรวม เก็บไว้เพื่อให้ย้อนดูที่มาของแต่ละแถวได้
ตัวอย่างข้อมูล (Example)
{
"request_text": "พี่ครับ รบกวนช่วยตรวจสอบสิทธิ์การเข้าใช้งานระบบบัญชีให้หน่อยครับ พอดีเข้าไปดูข้อมูลไม่ได้",
"category": "Account & Access",
"priority": "High",
"sentiment": "Neutral",
"_generation": "{'target_category': 'Account & Access', 'target_priority': 'High', 'target_sentiment': 'Neutral'}"
}
การกระจายของหมวดหมู่ (Label Distribution)
| หมวด (category) | จำนวน |
|---|---|
| Software & System | 350 |
| Hardware & Device | 340 |
| Information & Policy | 340 |
| Account & Access | 180 |
| Procurement Request | 170 |
| Network Issue | 170 |
| Data & Database | 170 |
| Meeting/Resource Request | 170 |
| Email & Communication | 170 |
| Quotation Request | 160 |
| Billing & Payment | 160 |
| HR & Employee Service | 160 |
| Security Issue | 160 |
| Plan/Budget Approval | 150 |
หมวดที่เกิดจากการรวมมีขนาดราว 2 เท่าของหมวดอื่น แนะนำให้ใช้ class weights หรือ stratified split ตอนเทรน
| priority | จำนวน | sentiment | จำนวน | |
|---|---|---|---|---|
| High | 999 | Neutral | 1,146 | |
| Medium | 997 | Negative | 998 | |
| Low | 854 | Positive | 706 |
การปรับจาก 18 เป็น 14 หมวด (Taxonomy Revision)
ชุดข้อมูลต้นฉบับมี 18 หมวด จำนวน 3,000 แถว ผลวิเคราะห์ error จากการเทรนโมเดลบน taxonomy เดิมพบปัญหา 2 แบบ จึงปรับดังนี้
1. ตัดหมวด Other (150 แถว) เพราะหมวดนี้นิยามด้วยการปฏิเสธ ("ไม่เข้าหมวดอื่น") ทำให้โมเดลเรียนรู้รูปแบบได้ยาก
2. รวม 3 คู่หมวดที่สับสนกันแบบสองทาง ซึ่งสะท้อนว่าขอบเขตนิยามทับซ้อนกัน
| หมวดเดิม | หมวดใหม่ |
|---|---|
| Hardware Issue + Device & Peripheral | Hardware & Device |
| Software Issue + System Support | Software & System |
| General Information + Policy & Procedure | Information & Policy |
หมวดอื่นอีก 11 หมวดคงชื่อและข้อมูลเดิมไว้ทั้งหมด การปรับทำด้วยการ relabel ก่อนเทรน ไม่ได้ map ผลทำนายภายหลัง
วิธีใช้งาน (Usage)
from datasets import load_dataset
ds = load_dataset("ShinRunner1st/ServiceHub_Dataset_14Cat")
print(ds["train"][0])
แบ่ง train/test แบบ stratified:
ds = ds["train"].class_encode_column("category")
split = ds.train_test_split(test_size=0.2, stratify_by_column="category", seed=2026)
ข้อจำกัด (Limitations)
- Synthetic data: ข้อความทั้งหมดสร้างด้วย LLM จึงอาจมีรูปแบบภาษาที่สม่ำเสมอกว่าคำร้องจริง ผลลัพธ์ของโมเดลที่เทรนบนชุดนี้อาจไม่สะท้อนประสิทธิภาพบนข้อมูลจริงในองค์กร
- Label มาจาก prompt:
category,priorityและsentimentคือค่าที่สั่งให้ LLM สร้าง ไม่ใช่การแท็กโดยมนุษย์ ข้อความบางแถวอาจไม่ตรงกับ label ที่ได้รับ (label noise) - Class imbalance: หมวดที่รวมกันมีขนาดใหญ่กว่าหมวดอื่นประมาณ 2 เท่า
- ขอบเขตโดเมน: ครอบคลุมคำร้องด้าน IT และงานสำนักงานทั่วไปเท่านั้น
- ไม่มีข้อมูลส่วนบุคคลจริง: เนื่องจากเป็นข้อมูลสังเคราะห์ ไม่ได้มาจากคำร้องของผู้ใช้จริง
License
เผยแพร่ภายใต้ CC BY 4.0
Citation
@misc{servicethub_thai_14cat,
title = {ServiceThub Thai IT Service Requests (14 categories)},
author = {ShinRunner1st},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/ShinRunner1st/ServiceHub_Dataset_14Cat}}
}