pretty_name: LexBench-SimpleQA-Verified
license:
- apache-2.0
- mit
tags:
- benchmark
- evaluation
- webstone
- web-search
- information-retrieval
- retrieval-evaluation
- question-answering
- simpleqa
- factuality
task_categories:
- text-retrieval
source_datasets:
- extended|google/simpleqa-verified
language:
- en
configs:
- config_name: factual-v1
data_files:
- split: test
path: data/factual-v1.jsonl
LexBench-SimpleQA-Verified
Short name: LB-SQAV.
A 831-question retrieval edition of Google SimpleQA Verified, used by Webstone to compare search tools. It retains publisher questions, reference answers and supporting URLs so a search result list can be checked for reference-page retrieval and answer-bearing text.
Google DeepMind and Google Research designed the original benchmark to measure a model's factual knowledge without search tools. This derivative changes the task to retrieval evaluation. Its retrieval metrics are not official SimpleQA Verified accuracy or evidence of a model's closed-book factual knowledge.
| Viewer group | Source suite | Tasks | Scope |
|---|---|---|---|
factual-v1 |
factual-v1 |
831 | Unchanged publisher questions of at most 128 Unicode code points, with reference answers and URLs |
This dataset owns the factual retrieval inputs used alongside LexBench-Web-Search. Web Search references these definitions through its composition descriptor; it does not publish another copy of the questions or source receipt.
Selection and changes
The factual-v1 configuration contains every question in the pinned 1,000-row publisher CSV whose problem has at most 128 Unicode code points, measured by Python's len() on the decoded string. This selects 831 questions; longer questions are excluded before collection. There is no truncation, rewriting or selection based on search results.
Publisher row order, question text and reference answer text are retained. Case IDs use sqv- followed by the publisher's original_index, preserving the link to OpenAI SimpleQA. Supporting URLs are split from the CSV field, deduplicated in order, and stripped of unmatched trailing closing parentheses. Topic and answer-type labels are retained. The publisher's multi_step and requires_reasoning fields are not part of these derived task records.
The source manifest records the exact publisher CSV revision and SHA-256, selection rule and expected counts. The selection implementation rebuilds the population from that input. This release contains no search-provider rankings, relevance grades or measured results.
Data and schema
| Resource | Contents |
|---|---|
data/factual-v1.jsonl |
Dataset-viewer rows: id, input, family, language, task |
definitions/datasets/factual-v1.json |
Complete task records and source-reference bindings |
definitions/factual-source.json |
Upstream input and deterministic selection contract |
manifest.json |
Release source commit, file sizes and SHA-256 hashes |
The viewer's input is the unchanged question. Parse task as JSON for id, query, reference_answer, reference_urls, topic, answer_type, provenance and derivation. Records have family: web, category: factual and language: en.
reference_urls retains the publisher's URLs. equivalent_reference_urls adds independently verified copies of those documents. Candidates come from fixed mirror URL templates applied to the reference URLs, without consulting search results. A copy needs the same article title and at least 80% matching characters in its normalized lead text, in contiguous blocks of at least 12 characters. Verification uses up to 4,000 lead characters and requires at least 200 normalized characters from each page. Failed requests and short pages establish no equivalence. equivalent_reference_provenance records the reference URL, verification method, observation date and source hashes for each accepted copy. The dataset's reference_equivalence record fixes these rules; url_equivalence selects the generic document-URL policy.
Load and evaluate
Choose a full Hugging Face commit from this dataset's revision history and retain it with your evaluation:
import json
from datasets import load_dataset
rows = load_dataset(
"Lexmount/LexBench-SimpleQA-Verified",
"factual-v1",
split="test",
revision="YOUR_FULL_HF_COMMIT",
)
task = json.loads(rows[0]["task"])
query = task["query"]
answer, supporting_urls = task["reference_answer"], task["reference_urls"]
definitions.tar.gz contains the same definition bytes as definitions/. Webstone's data lock selects the installed snapshot. Follow the Web Search guide to declare a profile, collect provider responses and build a report with saved evidence. The certified collector is scripts/run_certified_benchmark.py web-search.
Reference Hit@k tests whether a supplied supporting URL appears in the first k results; MRR@k measures the first such hit's reciprocal rank. Answer Match@k checks answer cues in returned text. Source references are not an exhaustive set of relevant pages, and lexical answer matches do not establish that a result correctly answers the question. Report these measures separately from graded relevance. Searchable copies of the dataset can also create contamination; Webstone exposes a separate contamination diagnostic while retaining the original results.
License and citations
The Google dataset card declares MIT. The derived definitions retain attribution to Google DeepMind / Google Research and to the original OpenAI SimpleQA, copyright 2024 OpenAI. Project-authored additions use Apache-2.0. See LICENSE, THIRD_PARTY.md and the retained notices. Supporting website URLs do not grant a license to redistribute the referenced pages.
Cite both upstream works when using this derivative, and identify the selected dataset revision:
Haas, Yona, D'Antonio, Goldshtein and Das (2025). SimpleQA Verified: A Reliable Factuality Benchmark to Measure Parametric Knowledge.
Wei et al. (2024). Measuring Short-Form Factuality in Large Language Models. OpenAI SimpleQA.
For the harness, cite Webstone: Benchmarking Browsers and Web Tools for AI Agents.
LexBench-Web-Search contains the separate scenario-query suites. LexBench-Web-Fetch and LexBench-WebMainBench evaluate fetching and extraction with their own task populations.