Yuvalos/sweet-spec-offers

SweetSpec - Synthetic B2B Sales Offers Dataset Overview

SweetSpec is a fully synthetic dataset consisting of 10,000 B2B sales offers, created for educational and experimental purposes. The dataset is designed to support the development and evaluation of AI-powered systems such as semantic similarity search, recommendation engines, and pricing-related machine learning workflows.

All data included in this dataset is fictional and privacy-safe. No real customers, products, pricing logic, or proprietary business information are used.

The dataset was generated using a hybrid synthetic data generation approach that combines a Hugging Face pretrained language model with structured templates. This approach ensures natural-language realism while maintaining a consistent and well-defined schema suitable for downstream analysis.

Dataset Structure

Each row in the dataset represents a single synthetic sales offer and includes the following fields:

offer_id: Synthetic unique identifier for the offer

fake_client: Fictional client name

region: Target market region

customer_type: Customer type (private label, distributor, pharma, ecommerce)

product_type: Product category

dietary_tags: Dietary features associated with the product

certifications: Product certifications

flavors: Flavor variants

moq_kg: Minimum order quantity in kilograms

lead_time_days: Estimated lead time in days

packaging: Packaging format

request_text: Free-text buyer request

offer_text: Free-text seller response

synthetic_price_per_kg: Synthetic price label used for modeling and analysis

Exploratory Data Analysis (EDA) Summary

The dataset contains exactly 10,000 synthetic sales offers, with no missing values across fields.

Product Type Distribution

Product categories are evenly distributed across all product types, avoiding class imbalance and supporting fair evaluation for similarity-based retrieval tasks.

Product Type Distribution Plot

Synthetic Price Distribution

Synthetic prices follow a smooth and realistic distribution, centered around a reasonable mean with natural variance. This distribution supports use cases such as pricing prediction and benchmarking.

Synthetic Price Distribution Plot

Price vs. Minimum Order Quantity (MOQ)

A weak negative relationship is observed between minimum order quantity (MOQ) and price per kilogram. This reflects plausible volume-discount behavior without enforcing deterministic pricing rules.

Price vs MOQ Scatter Plot

Offer Text Length Distribution

Seller offer texts exhibit consistent length across the dataset, making the data suitable for embedding-based semantic similarity search and other natural language processing tasks.

Offer Text Length Distribution Plot

Region Distribution

Regional coverage is balanced across multiple geographic markets, ensuring diversity and robustness in downstream machine learning workflows.

Region Distribution Plot

Overall, the exploratory data analysis confirms that the dataset is well-structured, diverse, and realistic, and is appropriate for AI applications such as semantic search, recommendation systems, and predictive modeling.

Files Included

sweet_spec_offers.csv — Main dataset in CSV format

sweet_spec_offers.parquet — Main dataset in Parquet format (recommended for performance)

Final_Project_ipynb_Yuval_Malka.ipynb — Complete synthetic data generation and EDA notebook

Disclaimer

This dataset is fully synthetic and was generated exclusively for educational and demonstration purposes. It does not contain real customer data, real pricing, real products, or proprietary business information.