tkdonda/indian-fashion-apparel-instruct

Indian Fashion & Apparel Product Data — Instruction Tuning Dataset

Dataset Description

A large-scale, instruction-tuning dataset covering 25 Indian fashion and apparel categories with 5,000 records per category (125,000 total entries). Designed for fine-tuning LLMs to generate product descriptions, extract product attributes, create SEO titles, and power AI-driven e-commerce tools for Indian fashion brands and platforms.

• Repository: tkdonda/indian-fashion-apparel-instruct
• Author: Tushar Donda
• License: CC BY 4.0
• Language: English
• Generated with: Google Gemini

---

Dataset Summary

India's e-commerce market is growing at 18.7% CAGR, reaching $326 billion by 2029. Yet most AI training datasets for fashion are biased toward Western clothing. This dataset fills that gap by providing rich, structured instruction-response pairs for Indian ethnic and contemporary apparel — from sarees and lehengas to sherwanis and juttis.

Each entry is formatted in the Alpaca instruction-tuning format (instruction, input, output), making it directly compatible with frameworks like Axolotl, LLaMA-Factory, Unsloth, and Hugging Face TRL.

---

Supported Tasks

Task	Description
Product Description Generation	Generate detailed, attractive product descriptions for Indian fashion items
SEO Title Optimization	Rewrite product titles for better marketplace visibility (Amazon, Flipkart, Myntra)
Attribute Extraction	Extract structured key-value attributes from unstructured product notes
Product Comparison	Compare two products and recommend based on occasion or user need
Outfit & Styling Suggestions	Suggest complete outfit pairings and styling tips
Bullet-Point Highlights	Create 4-point product highlights for app cards and quick-view panels
Care Instructions	Generate fabric-appropriate care and maintenance instructions

---

Categories Covered

The dataset is split into 25 individual JSONL files — one per category — each containing 5,000 records:

Category	File
Anarkali Dress	data/anarkali_dress-train.jsonl
Indo-Western Dress	data/indo_western_dress-train.jsonl
Lehenga Choli	data/lehenga_choli-train.jsonl
Men's Casual Shoes	data/men_casual_shoes-train.jsonl
Men's Formal Shoes	data/men_formal_shoes-train.jsonl
Men's Jeans	data/men_jeans-train.jsonl
Men's Kurta	data/men_kurta-train.jsonl
Men's Sherwani	data/men_sherwani-train.jsonl
Men's Shirt	data/men_shirt-train.jsonl
Men's Trousers	data/men_trousers-train.jsonl
Men's T-Shirt	data/men_tshirt-train.jsonl
Salwar Kameez	data/salwar_kameez-train.jsonl
Saree	data/saree-train.jsonl
Unisex Sneakers	data/unisex_sneakers-train.jsonl
Women's Dress	data/women_dress-train.jsonl
Women's Handbag	data/women_handbag-train.jsonl
Women's Heels	data/women_heels-train.jsonl
Women's Jutti / Mojari	data/women_jutti_mojari-train.jsonl
Women's Kurta	data/women_kurta-train.jsonl
Women's Sandals	data/women_sandals-train.jsonl
Women's Shirt	data/women_shirt-train.jsonl
Women's Skirt	data/women_skirt-train.jsonl
Women's Top	data/women_top-train.jsonl
Women's Trousers	data/women_trousers-train.jsonl
Women's T-Shirt	data/women_tshirt-train.jsonl

---

Dataset Structure

Data Format

Each record follows the Alpaca instruction-tuning format:

{
  "instruction": "Write a detailed and attractive product description for the following Anarkali dress.",
  "input": "Category: anarkali_dress\nType: palazzo Anarkali\nSilhouette: layered tiered flare\nBottom: sharara\nDupatta: with chiffon dupatta\nFabric: embroidered net\nEmbellishment: gota patti work\nNeckline: jewel neck\nSleeve: cold shoulder\nLength: floor length (maxi)\nFit: slim fit\nColor: navy blue\nSet includes: single piece Anarkali gown\nOccasion: sangeet",
  "output": "Radiate timeless elegance in this stunning navy blue palazzo Anarkali crafted in luxurious embroidered net. The layered tiered flare creates a graceful, flowing look that is perfect for sangeet. Featuring intricate gota patti work detailing, a flattering jewel neck neckline, and cold shoulder sleeves, this floor length (maxi) silhouette is both regal and comfortable. Comes with chiffon dupatta and paired with a sharara — single piece Anarkali gown. The slim fit ensures a flattering drape for all body types. Care: hand wash separately; iron on reverse side on low heat."
}

Data Fields

Field	Type	Description
instruction	string	The task prompt given to the model
input	string	Structured product attributes as key-value pairs
output	string	The expected model response (description, title, attributes, etc.)

Key Input Attributes (vary by category)

• Apparel: Category, Type, Fabric, Embellishment, Neckline, Sleeve, Length, Fit, Color, Occasion, Set Includes
• Footwear: Type, Style, Toe Shape, Upper Material, Sole, Closure, Insole, Detailing, Occasion
• Accessories: Type, Style, Size, Compartment, Strap, Outer Material, Hardware, Embellishment, Occasion

---

How to Load

Load a Single Category

from datasets import load_dataset

# Load only saree data
ds = load_dataset(
    "tkdonda/indian-fashion-apparel-instruct",
    data_files="data/saree-train.jsonl",
    split="train"
)
print(ds)

Load All Categories

from datasets import load_dataset

ds = load_dataset(
    "tkdonda/indian-fashion-apparel-instruct",
    split="train"
)
print(f"Total records: {len(ds)}")

Load Multiple Specific Categories

from datasets import load_dataset

ds = load_dataset(
    "tkdonda/indian-fashion-apparel-instruct",
    data_files={
        "saree": "data/saree-train.jsonl",
        "lehenga_choli": "data/lehenga_choli-train.jsonl",
        "anarkali_dress": "data/anarkali_dress-train.jsonl"
    }
)

---

Use Cases

1. Fine-tune an LLM for Product Description Generation

Fine-tune models like LLaMA 3, Mistral, or Gemma to auto-generate product listings for Indian fashion brands. Ideal for platforms like Meesho, Myntra, Flipkart Fashion, or D2C Indian brand websites.

2. Build a Product Information Management (PIM) Tool

Use the attribute extraction entries to train a model that parses unstructured product notes into clean, structured data — critical for managing multichannel listings.

3. Train an SEO Title Optimizer

Fine-tune a model on the SEO title entries to auto-generate marketplace-optimized titles with attributes in the right order (Color → Material → Type → Key Feature → Occasion).

4. Power an AI Fashion Stylist Chatbot

Use the outfit suggestion and product comparison entries to build a conversational stylist that recommends what to wear for specific occasions (sangeet, Navratri, office, beach resort).

5. Train Recommendation & Comparison Models

Product comparison entries include occasion-based recommendations — useful for building "best for your occasion" recommendation engines on Indian e-commerce platforms.

---

Dataset Statistics

Metric	Value
Total Records	~125,000
Categories	25
Records per Category	~5,000
Instruction Task Types	7
Format	Alpaca (JSONL)
Language	English
Domain	Indian Fashion & Apparel
Generation Model	Google Gemini

---

Dataset Creation

Curation Rationale

Existing fashion datasets (IndoFashion, DeepFashion, Fashionpedia) focus primarily on image classification for Western clothing. There is a significant gap in text-based, instruction-tuned datasets for Indian ethnic and contemporary apparel. This dataset was created to serve Indian AI developers, researchers, and e-commerce companies building NLP tools for the Indian market.

Source Data

All product attribute schemas were curated by the dataset author based on real Indian e-commerce product listing structures observed on platforms like Myntra, Flipkart, Amazon India, and Meesho. Product descriptions and outputs were generated using Google Gemini with carefully crafted prompts to ensure cultural accuracy, garment terminology authenticity, and output quality.

Annotations

• Annotation Type: Machine-generated via Google Gemini
• Quality Control: Outputs reviewed for Indian garment terminology accuracy (gota patti, bandhani, resham embroidery, jodhpuri pants, meenakari, etc.), care instruction completeness, and structural consistency

Personal and Sensitive Information

This dataset contains no personal information, user data, or real customer records. All product data is synthetically structured.

---

Considerations for Using the Data

Social Impact

This dataset supports the democratization of AI tools for Indian e-commerce — particularly for small and mid-size Indian brands (Tier 2 and Tier 3 cities) that cannot afford enterprise PIM or AI writing tools.

Discussion of Biases

• Outputs reflect contemporary Indian e-commerce writing conventions and may carry stylistic biases toward aspirational language ("Command every room," "Radiate timeless elegance")
• Occasion tags reflect Indian cultural events (sangeet, Navratri, puja) which may not generalize to non-Indian contexts
• Care instructions are general and should not replace manufacturer-specific guidance

Known Limitations

• All text is in English only — does not include Hindi, Gujarati, or other Indian regional languages
• Brand names are not included (generic brand references only)
• No product images are included — this is a text-only dataset

---

Additional Information

Related Datasets by the Same Author

• 🔗 tkdonda/gujarati-instruct-alpaca — Gujarati language instruction-tuning dataset

Licensing Information

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

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material for any purpose, including commercial use

Under the condition that you give appropriate credit to the author (Tushar Donda / tkdonda).

Citation

If you use this dataset in your research or product, please cite it as:

@dataset{donda2025indianfashion,
  author    = {Tushar Donda},
  title     = {Indian Fashion \& Apparel Product Data — Instruction Tuning Dataset},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/tkdonda/indian-fashion-apparel-instruct}
}

Contributions

Dataset created and curated by Tushar Donda (@tkdonda). If you find issues or want to contribute additional categories, please open a Discussion on this dataset page.