neurocheckout-ai/synthetic-product-recommendation-examples

Synthetic Product Recommendation Examples

An entirely synthetic bilingual dataset of ecommerce discovery queries paired with candidate products and graded relevance judgments. It is designed for educational retrieval, reranking, and recommendation experiments and contains no private catalog, merchant, customer, behavioral, or transaction data.

Dataset Description

The dataset contains twenty English and French queries with ten candidates per query. It complements NeuroCheckout's public Product Similarity Demo, but its relevance grades are controlled synthetic annotations, not cosine-similarity outputs from a model.

Each query has this candidate distribution:

Grade	Label	Candidates per query	Interpretation
3	strong_semantic_match	2	Requested product type and all three core attributes match
2	related_products	2	Product category and two core attributes match
1	weak_semantic_relation	2	One attribute matches, but the product type is adjacent
0	different_products	4	Different category with no core attribute match

These labels align conceptually with the four educational bands in the similarity Space. They are relevance judgments for evaluating a ranking, not probabilities or universal merchandising decisions.

Composition

Dimension	Count
Total pairs	200
Unique queries	20
Candidates per query	10
Train	120
Validation	40
Test	40
English pairs	100
French pairs	100
Grade 3 pairs	40
Grade 2 pairs	40
Grade 1 pairs	40
Grade 0 pairs	80

Ten fictional product families are used. Both language variants and all candidate products from one family remain in the same split. Candidate product identifiers never cross split boundaries.

Schema

Field	Description
id	Stable pair identifier
query_id, query_text, query_language	Synthetic discovery query information
target_category	Requested product category
candidate_product_id	Stable fictional catalog identifier
candidate_title, candidate_description, candidate_language	Synthetic localized candidate content
candidate_category	Candidate product category
candidate_attributes	Controlled synthetic attributes present on the candidate
matching_query_attributes	Requested core attributes present on the candidate
missing_query_attributes	Requested core attributes absent from the candidate
relevance_grade	Integer relevance judgment from zero to three
relevance_label	Human-readable relevance band
pair_type	Exact target, alternative, adjacent hard negative, or cross-category negative
relevance_rationale	Template explanation of the grade
synthetic	Always true

Usage

from datasets import load_dataset

dataset = load_dataset("neurocheckout-ai/synthetic-product-recommendation-examples")
test_pairs = dataset["test"]

Group rows by query_id, rank the ten candidates with a lexical, embedding, or reranking approach, and use relevance_grade as the qrel. Suitable educational metrics include Recall@K, mean reciprocal rank, and NDCG@10.

Generation and Validation

The committed generate_dataset.py script creates the JSON Lines files deterministically. Relevance is assigned from controlled category membership and attribute intersections rather than model output. The script verifies:

• exact split, language, query, and grade counts;
• ten unique candidate pairs per query with the intended grade distribution;
• consistency between grades, labels, pair types, matching attributes, and missing attributes;
• unique pair identifiers and query-product combinations;
• isolation of every candidate product identifier to one split;
• absence of email addresses, phone-like strings, and long digit sequences.

Limitations and Responsible Use

• The catalog, queries, attributes, and relevance labels are synthetic and deliberately structured.
• The candidate set is small and balanced for teaching; real retrieval systems operate over much larger, changing catalogs.
• Attribute overlap makes the task easier than noisy natural-language shopping behavior.
• The dataset does not include clicks, purchases, popularity, inventory, price, margin, user identity, or causal preference signals.
• Relevance can depend on context, availability, compatibility, accessibility, safety, and merchandising constraints not represented here.
• Results on this dataset do not establish production quality, fairness, safety, business impact, or regulatory compliance.
• The examples do not reveal or claim to represent NeuroCheckout's internal catalog, models, customers, or recommendation logic.

Do not use the dataset to profile people, infer sensitive characteristics, manipulate vulnerable users, or claim guaranteed product relevance. Production systems should measure long-tail coverage and popularity bias, enforce catalog rules separately, and provide user control.

License

The dataset and generation script are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

Learn More

Read NeuroCheckout's educational guide: AI Models for Ecommerce.