Dataset Card for items_lite

I4031/items_lite is a standardized, LLM-preprocessed product dataset created for "The Price is Right" capstone project in LLM Engineering.

Building upon I4031/items_raw_lite, this dataset transforms noisy raw e-commerce metadata into clean, structured, and concise product summaries generated by an LLM (gpt-oss-20b). This standard format strips uninformative part numbers and boilerplate, making it optimal for downstream price regression, RAG indexing, and fine-tuning frontier models.

---

Dataset Summary

• Repository: I4031/items_lite
• Total Datapoints: 22,000 products
• Splits:
  • train: 20,000 items
  • validation: 1,000 items
  • test: 1,000 items
• Price Range: $0.50 – $999.49 USD
• Source Data: McAuley-Lab Amazon Reviews 2023 (curated via I4031/items_raw_lite)
• Preprocessing Model: openai/gpt-oss-20b (batch rewritten via Groq / LiteLLM)

---

Dataset Structure

Each datapoint contains the following fields:

Feature	Type	Description
title	string	Original product title
category	string	Amazon category (e.g., Tools_and_Home_Improvement, Appliances, Electronics)
price	float	Actual price in USD (ground-truth target)
summary	string	Standardized LLM-generated product summary (5-line structured text)
weight	float	Item shipping weight normalized to pounds (if available, else 0.0)
prompt	string / null	Optional formatted prompt string for fine-tuning

Summary Format

The summary field follows a strict 5-line schema designed to maximize signal and remove noisy identifiers:

Title: Rewritten short precise title
Category: Category name
Brand: Brand name
Description: 1 sentence product description
Details: 1 sentence on features and materials

---

Data Pipeline & Preprocessing

1. Raw Curation (Day 1):
   • Filtered from the McAuley-Lab/Amazon-Reviews-2023 dataset.
   • Cleaned noise (best seller ranks, uninformative metadata keys, serial codes).
   • Price range clamped to $0.50 – $999.49 with quadratic price-rebalancing to ensure strong representation of higher-priced goods.
2. LLM Standardization (Day 2):
   • Every raw product description (full) was processed through openai/gpt-oss-20b with reasoning effort set to low. The processing was performed locally using Ollama and RTX 5070Ti.
   • Stripped out serial numbers, ASINs, and part numbers so models learn to predict price based on intrinsic product features, brand, and build quality rather than memorized identifiers.
   • Raw full and index id fields were cleared prior to Hub upload to minimize parquet footprint.

---

Quickstart Usage

Using Hugging Face datasets

from datasets import load_dataset

# Load dataset
ds = load_dataset("I4031/items_lite")

# Inspect a train record
sample = ds["train"][0]
print("Category:", sample["category"])
print("Price:   $", sample["price"])
print("\nProduct Summary:\n" + sample["summary"])

Using Course pricer Module

from pricer.items import Item

# Load directly into typed Item objects
train, val, test = Item.from_hub("I4031/items_lite")

print(f"Loaded {len(train):,} train, {len(val):,} val, {len(test):,} test items")
print(train[0].summary)
print(f"Target: ${train[0].price}")

---

Intended Uses & Applications

• Price Prediction: Regression models estimating retail prices from textual product features.
• LLM Fine-Tuning: Training models (e.g., Llama 3.2 3B, GPT-4o-mini) to reason about value and product pricing.
• Vector Search / RAG: Building semantic embeddings and product retrieval vector stores.
• Zero-Shot & Few-Shot Evaluation: Benchmarking frontier reasoning models on structured tabular and numerical extraction tasks.

---

Source & Attribution

• Built as part of the LLM Engineering course capstone project ("The Price is Right").
• Base data derived from the Amazon-Reviews-2023 dataset curated by McAuley Lab at UC San Diego.

---

License & Usage Terms

• License: Academic Non-Commercial / Research use.
• Terms: Subject to Amazon's Conditions of Use and the source dataset terms. Not intended for commercial applications.