esci-product-search-pairs

Ready-to-train contrastive pairs for e-commerce product-search embedding models, built from the real Amazon ESCI Shopping Queries Dataset, with explicitly-labeled synthetic noisy-query augmentation on top. This is the exact dataset used to train ecommerce-product-search-embeddings.

Every row: query, positive, negative_1..3 — a real (or labeled-synthetic) shopping query, a product title with a human "Exact" relevance judgment, and three hard negative titles drawn from the same query's human-judged pool (Irrelevant > Complement > Substitute priority), i.e. the free graded hard negatives that make ESCI special.

What is real and what is synthetic?

Every row carries a source column. No synthetic row is ever presented as real:

source	rows	what it is
real_esci	115,631	real query, real human-judged positive + negatives, straight from ESCI
synthetic_typo	16,914	rule-based keyboard typos injected into a real parent query (seeded, reproducible)
synthetic_abbrev	2,112	rule-based shopping abbreviations/compressions of a real parent query ("christmas"→"xmas", stopword drops)
synthetic_intent	1,211	LLM-generated intent rephrasing of a real parent query ("nonstick pan" → "pan that eggs dont stick to"), each one adversarially verified by an independent LLM judge for intent faithfulness before inclusion (10% of candidates were rejected)

Synthetic rows total 14.9% of the dataset. Each keeps parent_query (the real query it derives from) — and inherits the parent's human-judged positive and negatives, so labels are always real human judgments; only the query surface form is synthetic.

Where does the real data come from?

The Amazon ESCI Shopping Queries Dataset (Apache-2.0; Reddy et al. 2022): real anonymized Amazon search queries with graded human relevance labels (Exact / Substitute / Complement / Irrelevant). We use the US locale, task-1 (small_version) TRAIN split only — 427,655 judgments over 20,888 queries, pinned at mirror revision 8113b17a. The official ESCI test split is not in this dataset in any form; it stays clean for evaluation. We verified zero normalized-query-text overlap between this data and the official test split.

A 1,500-query dev slice (dev_pairs split, by query id) is included for checkpoint selection.

How do I train an embedding model on this?

from datasets import load_dataset
from sentence_transformers import SentenceTransformer, SentenceTransformerTrainer, SentenceTransformerTrainingArguments
from sentence_transformers.losses import MultipleNegativesRankingLoss
from sentence_transformers.training_args import BatchSamplers

ds = load_dataset("albertobarnabo/esci-product-search-pairs", split="train")
ds = ds.select_columns(["query", "positive", "negative_1", "negative_2", "negative_3"])

model = SentenceTransformer("BAAI/bge-small-en-v1.5")
trainer = SentenceTransformerTrainer(
    model=model,
    train_dataset=ds,
    loss=MultipleNegativesRankingLoss(model),
    args=SentenceTransformerTrainingArguments(
        output_dir="out", num_train_epochs=1, per_device_train_batch_size=64,
        batch_sampler=BatchSamplers.NO_DUPLICATES,
    ),
)
trainer.train()

Want only the real rows? ds = ds.filter(lambda r: r["source"] == "real_esci").

Schema

column	type	notes
query	str	shopping query (real, or synthetic variant — see source)
positive	str	product title, human-labeled Exact for the (parent) query
negative_1..3	str	hard negatives from the same query's judged pool
negative_1..3_kind	str	Irrelevant / Complement / Substitute / random (pad)
source	str	real_esci / synthetic_typo / synthetic_abbrev / synthetic_intent
parent_query	str	the real query a synthetic variant derives from (empty for real rows)
query_id	int	ESCI query id (shared by a parent and its variants)
positive_product_id	str	ESCI/Amazon product id (ASIN)

Citation

Cite the underlying ESCI dataset:

@article{reddy2022shopping,
  title={Shopping Queries Dataset: A Large-Scale {ESCI} Benchmark for Improving Product Search},
  author={Reddy, Chandan K. and M{\`a}rquez, Llu{\'i}s and Valero, Fran and Rao, Nikhil and Zaragoza, Hugo and Bandyopadhyay, Sambaran and Biswas, Arnab and Xing, Anlu and Subbian, Karthik},
  journal={arXiv preprint arXiv:2206.06588},
  year={2022}
}

Family: fine-tuned model · static ~30MB model · live demo.