Lezh1n/ecommerce-product-classification-by-categories

🛍️ E-Commerce Product Classification Dataset

Dataset Description

This dataset contains 31,851 realistic English product descriptions across 33 e-commerce categories, designed for training product classification models.

Key Features

• ✅ 31,851 samples with natural language variations
• ✅ 33 product categories covering major e-commerce segments
• ✅ 4 seller personas: Professional, Individual, Reseller, Minimal
• ✅ Realistic noise: Typos, abbreviations, casual language
• ✅ Length variation: 8-800 characters (2-120 words)
• ✅ English-only for international e-commerce applications

Dataset Statistics

Split	Samples	Percentage
Train	~22,295	70%
Validation	~4,777	15%
Test	~4,777	15%

Text Length:

• Min: 8 chars
• Max: 800 chars
• Mean: 291 chars
• Median: 270 chars

Word Count:

• Min: 2 words
• Max: 120 words
• Mean: 46.7 words
• Median: 46 words

Categories

The dataset covers 33 major e-commerce categories:

'arts_crafts', 'automotive_motorcycle', 'baby_maternity', 'bags_luggage', 'beauty_personal_care', 'bedding_bath', 'books_media', 'computers_networking', 'electronics', 'fashion_accessories', 'fashion_clothing', 'garden_outdoor_living', 'grocery_food', 'health_wellness', 'home_decor_lighting', 'home_furniture', 'industrial_commercial', 'jewelry', 'kitchen_dining', 'large_appliances', 'mobile_phones_tablets', 'musical_instruments', 'pet_supplies', 'shoes_footwear', 'small_appliances', 'software_digital_goods', 'sports_outdoors', 'stationery_office_supplies', 'tools_hardware', 'toys_games', 'video_games_gaming', 'watches'

Category Distribution (Top 10)

Category	Samples
home_furniture	1,000
toys_games	1,000
bags_luggage	1,000
beauty_personal_care	1,000
arts_crafts	999
large_appliances	999
pet_supplies	999
health_wellness	999
baby_maternity	999
bedding_bath	999

Dataset Structure

Data Fields

• id: Unique sample identifier (int)
• text: Product description (string)
• label: Product category (string)

Example Samples

Professional Seller:

{
  "id": 1,
  "text": "Sony WH-1000XM5 Wireless Noise-Cancelling Headphones - Premium over-ear headphones featuring industry-leading noise cancellation with 8 microphones and Integrated Processor V1. 30-hour battery life with quick charging (3 min = 3 hours). Multipoint connection for pairing with two devices. Crystal-clear calls with precise voice pickup. Includes carrying case, audio cable, and USB-C charging cable.",
  "label": "electronics"
}

Individual Seller:

{
  "id": 2,
  "text": "sony wh-1000xm5 headphones barely used. got these 3 months ago but dont use them much. sound quality is amazing, noise cancelling works great. comes with case and cables. $250 or best offer",
  "label": "electronics"
}

Reseller:

{
  "id": 3,
  "text": "Sony WH-1000XM5 BRAND NEW SEALED!! LATEST MODEL!! Industry-leading NOISE CANCELLATION!! 30 HOUR battery!! FAST FREE SHIPPING!! Ships same day!! Includes FULL WARRANTY!! LIMITED STOCK!! LOWEST PRICE!!",
  "label": "electronics"
}

Minimal:

{
  "id": 4,
  "text": "sony wh-1000xm5 like new. noise cancelling headphones. $240",
  "label": "electronics"
}

Use Cases

Primary Use Case: Product Classification

Train models to automatically categorize e-commerce product listings:

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load dataset
dataset = load_dataset("Lezh1n/ecommerce-product-classification-by-categories")

# Train classifier
model = AutoModelForSequenceClassification.from_pretrained(
    "bert-base-uncased",
    num_labels=32
)

# Your training code here...

Other Applications

• Text Generation: Train models to generate realistic product descriptions
• E-commerce Search: Improve product search and recommendation systems
• Data Augmentation: Expand existing product catalogs
• Multi-language Models: Fine-tune for other languages

Data Collection

Generation Process

1. Prompting Strategy: Four distinct seller personas (professional, individual, reseller, minimal)
2. LLM Generation: Generated using Gemma 2B with temperature 0.85
3. Quality Filtering:
   • Length validation (60-800 characters)
   • Content relevance checks
   • Duplicate removal
4. Noise Injection: Realistic typos, abbreviations, and casual language
5. Validation: Manual review of samples

Data Quality

• ✅ 99.8% valid samples (31,801 / 31,851)
• ✅ Stratified splits maintain category distribution
• ✅ Deduplicated using text similarity
• ✅ Balanced across categories (~1,000 per category)

Limitations

• English only: Not suitable for multilingual classification
• Synthetic generation: May not capture all real-world variations
• Category scope: Limited to 32 predefined categories
• Seller personas: May not represent all marketplace types
• Noise samples: Very few (50 samples, 0.2%) for "none" category

Ethical Considerations

• No personal information: All product descriptions are synthetic
• No harmful content: Filtered for appropriate content
• No copyrighted material: Original generated descriptions
• Fair use: Intended for research and commercial ML applications

Citation

If you use this dataset, please cite:

@dataset{ecommerce-product-classification-by-categories,
  title = {E-Commerce Product Classification Dataset},
  author = {Lezh1n},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Lezh1n/ecommerce-product-classification-by-categories}
}

License

This dataset is released under CC BY 4.0 (Creative Commons Attribution 4.0 International).

You are free to:

• ✅ Share and redistribute
• ✅ Adapt and build upon
• ✅ Use commercially

Under the following terms:

• Attribution required

Dataset Card Authors

• Lezh1n