openbenchmarks/OB-Company-Websearch-Val

OB Company Websearch Validation

50 multi-constraint company-discovery questions with a frozen, hand-labelled reference set. This validation set is for developing and tuning against the OpenBenchmarks Multi Turn Company Search Benchmark. Scores on it are not leaderboard scores: the leaderboard is computed on a separate held-out set, and none of these questions appear in it.

The benchmark holds the research agent and budgets fixed while varying the web search provider. The questions apply to both conditions: search-only, where the agent can search and read result snippets, and search and fetch, where it can also fetch pages.

Dataset contents

Each question asks for the complete set of companies satisfying three or four simultaneous constraints. This release contains:

• 50 questions
• 21 three-constraint questions and 29 four-constraint questions
• 340 reference-set memberships across the questions
• 334 unique canonical companies

Each row contains:

• case_key: stable question identifier.
• question: natural-language company-discovery request shown to the agent.
• family: broad construction family such as investor, accelerator, or funding.
• constraint_count and constraints: number and text of the simultaneous conditions.
• predicates: normalized axes and labels used to construct the question.
• census_metadata: dataset-construction checks.
• source_metadata: construction coverage metadata.
• sort_order: stable display order.
• gold: reference companies, including canonical name, domain, aliases, and notes.

Dataset-level release metadata and the source content hash are retained in metadata.json.

Evaluation

A fixed agent is asked each question and must return the complete matching company set. Returned companies are resolved to canonical identities and compared deterministically with gold:

• a reference company returned by the agent is a true positive;
• an extra company is a false positive;
• a missed reference company is a false negative.

Precision, recall, F1, and exact-set accuracy are computed per agent run.

This repository contains questions and labels. It does not contain vendor responses or benchmark scores.

Gold construction

Questions were generated from intersections of attributes in a company snapshot, keeping only questions where every constraint narrows the answer set. Fifteen web search APIs were each run on every question in both conditions. The companies they returned, together with the snapshot's own answers, were checked against each constraint by an automated judge and then reviewed by hand. The reference set was frozen after every question was reviewed.

Reproducibility

The open runner and offline judge are available at openbenchmarks-labs/multi-turn-company-search. The benchmark methodology and current results are published at openbenchmarks.com/multi-turn-company-search.