Ionio-ai/ecommerce-search-extraction

Ionio E-commerce Search Query Extraction

Built with: simula — schema-driven synthetic data generation with auditable taxonomy lineage.

An English synthetic dataset for training and evaluating systems that translate natural-language shopping requests into narrow, atomic, database-queryable JSON. It contains 10,985 accepted examples from a 13,000-attempt generation run. No accepted rows were trimmed from this release.

Each example pairs a realistic typed or spoken shopper query with a variable-schema extraction. The corpus covers product identity, brands, sizing, appearance, material, price, ratings, shipping, availability, condition, compatibility, technical specifications, sorting, and exclusions across a broad range of retail verticals.

Dataset structure

There is one train split in data/train.jsonl.

Column	Type	Description
id	string	Stable generated-example identifier.
attempt_index	integer	Original generation attempt index.
query	string	Natural-language shopper request.
extraction	string	Canonical compact JSON encoding of the structured extraction. Parse with json.loads.
product_vertical_path	string	Slash-separated taxonomy lineage for the retail vertical.
shopper_intent_path	string	Slash-separated taxonomy lineage for shopper intent.
attribute_focus_path	string	Slash-separated taxonomy lineage for the dominant attributes.
query_complexity_path	string	Slash-separated taxonomy lineage for linguistic shape and complexity.
strategy_id	string	Sampling strategy used for the example.
meta_prompt	string	Per-example generation plan retained for auditability.
complexified	boolean	Whether the generation plan used the complexification path.
generator_model	string	Model identifier recorded during generation.
critic_verdicts	string	JSON-encoded critic/refinement audit trail.
output_format	string	Output format requested from the generator.
schema_valid	boolean	Whether the generated envelope passed configured schema validation.
accepted	boolean	Whether the example passed the generation critic. Always true in this release.
created_at	string	UTC creation timestamp.

extraction is stored as JSON text rather than an Arrow struct because the extraction schema is intentionally dynamic and a field can validly have different JSON types across examples. This form is robust in the Hugging Face viewer and is directly suitable as a structured-generation target.

import json
from datasets import load_dataset

ds = load_dataset("Ionio-ai/ecommerce-search-extraction", split="train")
example = ds[0]
query = example["query"]
extraction = json.loads(example["extraction"])

Corpus statistics

Visual overview

Query length	Extraction size

Sampling strategy mix	Product vertical coverage

Metric	Value
Generation attempts	13,000
Accepted rows published	10,985
Acceptance rate	84.5%
Unique query–extraction pairs	10,981
Exact duplicate pairs retained	4
Unique query strings	10,977
Complexified examples	3,300 (30.0%)
Sampling strategies	5
Generator	deepseek/deepseek-v4-flash:nitro

Query length

Words	Rows	Share
1–10	1,821	16.6%
11–30	2,378	21.6%
31–60	2,832	25.8%
61–100	2,677	24.4%
101+	1,277	11.6%

The median query has 45 words (mean 51.35); observed queries range from 1 to 288 words.

Extraction size

Leaf values	Rows	Share
2–5	1,304	11.9%
6–11	2,836	25.8%
12–20	4,109	37.4%
21–30	2,555	23.3%
31+	181	1.6%

The median extraction contains 15 leaf values (mean 14.95). The most frequent top-level keys are product_type, rating_min, color, condition, brand, price_max, material, price, free_shipping, and sort_by.

Strategy distribution

Strategy	Rows	Share
general	3,853	35.1%
directed_detailed	1,931	17.6%
browsing_inspiration	1,911	17.4%
value_deals	1,867	17.0%
negation_complex_expressions	1,423	13.0%

The four-factor taxonomy covers product vertical, shopper intent, attribute focus, and query complexity. An automated lineage-coverage evaluation on the derived 10,000-row evaluation subset covered every taxonomy node at every configured level.

Generation and quality controls

The corpus was generated with simula, a taxonomy-conditioned synthetic-data pipeline. A depth-4 taxonomy and five sampling strategies controlled the distribution. Each row retained its sampled lineage and generation plan. Candidate records were checked against the configured envelope schema and passed through an automated critic/refinement stage; this release contains only accepted rows.

The schema guarantees an object containing query and extraction. Because extraction is an open, variable-schema object, schema_valid=true does not independently prove that every inner field is semantically correct. Critic acceptance is model-based rather than human annotation.

Intended uses

• Fine-tuning structured-generation or semantic-parsing models for shopping search.
• Developing natural-language interfaces for catalogs, filters, and product databases.
• Evaluating extraction behavior across query lengths, retail verticals, exclusions, ambiguity, and multi-constraint requests.
• Bootstrapping field normalization and schema-design experiments.

Limitations and responsible use

• The data is fully synthetic and may contain implausible products, inconsistent normalization, hallucinated constraints, brand references, or critic false positives.
• It reflects the behavior and biases of its generator, critic, prompts, and taxonomy as of June 2026; it is not a sample of real customer traffic.
• Four exact duplicate query–extraction pairs are intentionally retained because this release includes every accepted row. Consumers can deduplicate if their use case requires it.
• Do not treat prices, availability, ratings, compatibility, certifications, or product claims as factual or current.
• Before production use, perform domain-specific human review, validate the target schema, and test performance on real, consented data representative of the deployment population.

Privacy

No customer logs or deliberately collected personal data were used; all rows are synthetic. As with any large generated corpus, consumers should still run their own privacy and memorization checks. The complete generation prompts and responses are not published beyond the per-row audit fields in this dataset.

The SHA-256 digest of data/train.jsonl at publication is 5204e106dedb1525602d1dd772d7f186f78b11ea3758720236c61f3e3121eac3.

License

No explicit data license was supplied for this release, so the Hub metadata is marked other. Contact Ionio-ai for licensing and redistribution terms.

Citation

@dataset{ionio_ecommerce_search_extraction_2026,
  title        = {Ionio E-commerce Search Query Extraction},
  author       = {{Ionio-ai}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Ionio-ai/ecommerce-search-extraction}
}