Datasets:

NagaYu
/

specbench

pretty_name: SpecBench v0 (synthetic, dev split)
license: other
license_name: all-rights-reserved-evaluation-only
license_link: https://github.com/NagaYu/specsure/blob/main/LICENSE
language:
  - en
  - ja
task_categories:
  - text-classification
tags:
  - fact-verification
  - e-commerce
  - shopping-agents
  - synthetic
size_categories:
  - n<1K
configs:
  - config_name: dev
    data_files: specbench-v0-dev.jsonl

SpecBench v0 — synthetic dev split

Code, verifier and documentation: https://github.com/NagaYu/specsure

Product claims of the kind an AI shopping agent makes, each labelled true (PASS) or erroneous (FAIL, with error type) against an attribute-level spec snapshot of a fictional product.

28 synthetic products (22 Japanese-market, 6 US-market; chargers, ink, batteries, earbuds, phone cases, cables, hair dryers, snacks, cereal, chocolate, drink mix, skincare, sunscreen, shampoo), 531 claims: 295 true and 236 erroneous.
Error types: E1 model/generation mix-up (22), E2 variant mix-up (44), E3 fabricated number or unit error (56), E4 fabricated compatibility (30), E5 bundle or pack-count error (22), E7 ingredient/allergen error (24), E8 regional-spec mix-up (38). E6 (stale price/stock) is not covered.
Severity per claim: S1 safety/legal, S2 invalidates the purchase, S3 degrades the experience.

Fields

Field	Description
`id`	claim id (`g-…` true claim, `m-…` generated error)
`product`	brand, category, region (JP/US), model number, GTIN, English name, family id
`claim`	`claim_type`, `attribute`, `value_json` (JSON-encoded value), `unit`, `polarity` (affirm/deny), `product_ref` as an agent might state it
`label`	`expected` (PASS/FAIL), `error_type` (E1–E8 or null), `severity`, `method` (how the error was generated)
`evidence`	snapshot id, attribute path, gold value (JSON), source URL (`synthetic://…`), capture time

How it was made

Products and specs were written by hand as fictional data. True claims are generated from the snapshots with surface variation (full-width model numbers, GTIN-14, Japanese/English names, unit conversions). Errors are generated deterministically (fixed seed) by swapping in values from sibling generations, variants or regional versions, shifting numbers, swapping units, and adding or denying compatibility, bundle items and ingredients. Each error records which gold value it contradicts.

Splits and leakage

This is the dev side of a time split (snapshots captured before 2026-09-15). Sibling variants across splits are intentional hard negatives. The test side and the handwritten unseen challenge set are not included.

Limitations

Synthetic. Brands and products are fictional; GTINs use GS1 restricted-circulation prefixes and do not identify real products. Nothing here describes a real product.
Circularity. Labels were generated with the same vocabulary and tolerances as the reference verifier, so that verifier scores near 100% here; use the data to compare systems, not to estimate real-world accuracy.
No real agent answers. Claims are structured, not free text from shopping agents.
Allergen rules are provisional. The mandatory-allergen lists behind some labels (JP 8 / US 9 items) are pending legal review. Not for allergy advice.

Licence

Copyright (c) 2026 NagaYu. All rights reserved; published for evaluation only — you may view and use the data locally to evaluate it. Commercial use, redistribution and publishing copies require prior written permission (see the LICENSE; contact via the author's portfolio: https://nagayu-portfolio.nagayu.workers.dev/en).