README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: eval
        path: data/eval-*
datasets:
  - bprateek/amazon_product_description
dataset_info:
  features:
    - name: Product Name
      dtype: string
    - name: Category
      dtype: string
    - name: Description
      dtype: string
    - name: Selling Price
      dtype: string
    - name: Product Specification
      dtype: string
    - name: Image
      dtype: string
  splits:
    - name: train
      num_bytes: 12542887
      num_examples: 23993
    - name: test
      num_bytes: 3499375
      num_examples: 6665
    - name: eval
      num_bytes: 1376174
      num_examples: 2666
  download_size: 6391314
  dataset_size: 17418436
license: apache-2.0
task_categories:
  - image-classification
  - image-to-text
language:
  - en
size_categories:
  - 10K<n<100K

Dataset Creation and Processing Overview

This dataset underwent a comprehensive process of loading, cleaning, processing, and preparing, incorporating a range of data manipulation and NLP techniques to optimize its utility for machine learning models, particularly in natural language processing.

Data Loading and Initial Cleaning

• Source: Loaded from the Hugging Face dataset repository bprateek/amazon_product_description.
• Conversion to Pandas DataFrame: For ease of data manipulation.
• Null Value Removal: Rows with null values in the 'About Product' column were discarded.

Data Cleaning and NLP Processing

• Sentence Extraction: 'About Product' descriptions were split into sentences, identifying common phrases.
• Emoji and Special Character Removal: A regex function removed these elements from the product descriptions.
• Common Phrase Elimination: A function was used to strip common phrases from each product description.
• Improving Writing Standards: Adjusted capitalization, punctuation, and replaced '&' with 'and' for better readability and formalization.

Sentence Similarity Analysis

• Model Application: The pre-trained Sentence Transformer model 'all-MiniLM-L6-v2' was used.
• Sentence Comparison: Identified the most similar sentence to each product name within the cleaned product descriptions.
• Integration of Results: Added the most similar sentences as a new column 'Most_Similar_Sentence'.

Dataset Refinement

• Column Selection: Retained relevant columns for final dataset.
• Image URL Processing: Split multiple image URLs into individual URLs, removing specific unwanted URLs.
• Column Renaming: Renamed 'Most_Similar_Sentence' to 'Description'.

Image Validation

• Image URL Validation: Implemented a function to verify the validity of each image URL.
• Filtering Valid Images: Retained only rows with valid image URLs.

Dataset Splitting for Machine Learning

• Creation of Train, Test, and Eval Sets: Used scikit-learn's train_test_split for dataset division.

Hugging Face Dataset Preparation and Publishing

• Conversion to Dataset Objects: Converted each Pandas DataFrame (train, test, eval) into Hugging Face Dataset objects.
• Dataset Dictionary Assembly: Aggregated all splits into a DatasetDict.
• Publishing to Hugging Face Hub: The dataset was named "amazon-products" and pushed to the Hub for community access.

Dataset Card Information

• Configs: The dataset is split into train, test, and eval configurations, with paths specified for each.
• Features: Includes fields for Product Name, Category, Description, Selling Price, Product Specification, and Image.
• Splits: Detailed information on the number of bytes and examples for each dataset split.
• Sizes: Download and total dataset size specifications are provided.

For further details or to contribute to enhancing the dataset card, please refer to the Hugging Face Dataset Card Contribution Guide.