thepian/product-query-benchmark

Product Query Benchmark

A quality-filtered product search benchmark derived from the Amazon ESCI dataset, enriched with brand metadata from EUIPO (EU trademark registry) and Wikidata. Secondly, queries from ESCII are extended with variants involving origins, exclusions, certifications.

Only Tier-1 brands are included — brands that have a confirmed EUIPO trademark registration, giving a stable, verifiable brand identity (brand_euipo_id) for each product.

Files

File	Rows	Description
products.parquet	~260k	Products with ≥1 Exact relevance judgment, with brand enrichment
examples.parquet	~504k	Query-product relevance pairs (E/S/C/I), query text inlined
brand_aliases.parquet	~21k	Multilingual brand name variants (30 European languages, Wikidata-sourced)
brand_examples.parquet	~269k	Query-brand relevance pairs derived by aggregating product labels

Schema

products.parquet

Column	Type	Description
product_id	string	ESCI product_id (matches original Amazon ESCI dataset)
product_title	string	English product title
product_description	string	Long prose description (NULL for ~64% of products)
product_bullet_point	string	Feature bullet points (NULL for ~51% of products)
product_brand	string	Raw brand string from ESCI
brand_euipo_id	string	Stable EU trademark ID — use as brand key
brand_country	string	ISO 3166-1 alpha-2 country code (NULL for ~74% of brands)
brand_sector	string	Coarse sector derived from EUIPO Nice classes (see table below)
euipo_nice_classes	string	JSON array of EUIPO Nice classification numbers, e.g. [3, 5]

examples.parquet

Column	Type	Description
query_id	string	Stable MD5-derived hex ID for grouping all pairs from one query
query	string	Raw search query text
product_id	string	Joins to products.product_id. Either solely implied or directly referenced
esci_label	string	Exact / Substitute / Complement / Irrelevant
product_brand	string	Raw brand string in search
origin	string	Country of origin in search
certification	string	Certification referenced in search
exclusions	string array	Exclusions referenced in search
split	string	train / test (ESCI's original split)

brand_aliases.parquet

Column	Type	Description
brand_euipo_id	string	Joins to products.brand_euipo_id
brand_name	string	Canonical brand name
alias	string	Alternate name (translated, abbreviated, legal variant, etc.)
language	string	ISO 639-1 language code; NULL = language-agnostic
source	string	wikidata / euipo / manual

brand_examples.parquet

Brand-level relevance derived by aggregating product labels. For each (query, brand) pair, brand_label is the highest-priority label among all products from that brand judged for that query: Exact > Substitute > Complement > Irrelevant.

Column	Type	Description
query_id	string	Joins to examples.query_id
query	string	Raw search query text
brand_euipo_id	string	Joins to products.brand_euipo_id
brand_label	string	Exact / Substitute / Complement / Irrelevant
brand_origin	string	eu / non-eu / unknown — derived from brand_country; ~26% of brands have country data
split	string	train / test (inherited from query's examples)

Sector Distribution

Brand sector is derived from EUIPO Nice Classification goods classes (1–34 take priority over service classes 35–45).

Sector	Brands	Products
electronics	3,138	85,890
clothing	1,204	37,992
bags_luggage	548	20,481
home_living	1,448	27,821
sports_toys	475	9,793
personal_care	1,817	46,875
office_media	656	15,935
food	594	12,076
beverages	166	2,747
hardware	1,527	36,321
jewelry	380	9,837
medical	326	7,538
vehicles	335	6,096
other	545	12,408
TOTAL	13,159	331,810

Usage

from datasets import load_dataset

# Pin to a specific version for reproducible training runs
ds = load_dataset("thepian/product-query-benchmark", revision="v1.0.0")

products = ds["products"].to_pandas()
examples = ds["examples"].to_pandas()
brand_aliases = ds["brand_aliases"].to_pandas()

Or load individual files:

import pandas as pd

products = pd.read_parquet("hf://datasets/thepian/product-query-benchmark/products.parquet")
examples = pd.read_parquet("hf://datasets/thepian/product-query-benchmark/examples.parquet")

Versioning

This dataset uses semantic versioning via git tags. Always load with revision= for reproducible results — the default (HEAD) may change between runs.

Version	Date	Notes
v1.0.0	2026-04-24	Initial release: 13,159 EUIPO-verified brands, US locale

Schema stability: brand_country coverage is ~26% in v1 (Wikidata-sourced only). Product-level category inference and NACE codes are planned for v2.0.0.

License

• Relevance judgments and product metadata: CC BY-NC 4.0 (inherited from Amazon ESCI)
• Brand enrichment (EUIPO, Wikidata): open data, compatible with CC BY-NC 4.0

Commercial use requires a separate license from Amazon for the underlying ESCI data.

Citation

If you use this dataset, please cite the original Amazon ESCI paper:

@article{reddy2022shopping,
  title={Shopping Queries Dataset: A Large-Scale {ESCI} Benchmark for Improving Product Search},
  author={Reddy, Chandan K and Halverson, Llana and Deshpande, Ohad and others},
  journal={arXiv preprint arXiv:2206.06588},
  year={2022}
}