task_categories:
- question-answering
language:
- en
tags:
- web-search
- search-api
- company-news
- factual-lookup
- agents
- tool-use
size_categories:
- n<1K
configs:
- config_name: company_news
data_files:
- split: eval
path: data/eval.jsonl
OB News Websearch
100 atomic company-news questions that evaluate web search providers on
a factual lookup task. The model, extract prompt, and runner stay fixed; the search provider is the variable under test.
This is the public eval set. 100 question/answer pairs. Rows carry the gold answer, so they are not contamination-free. Use them to inspect the format and to run a local harness.
Benchmark
We evaluate web search on recent company events: financing, hires, product launches, acquisitions, office expansions, and headcount cuts.
The searcher receives one natural-language question and may return at most 10 hits. There is no query rewrite and no page fetch. An extract model reads the hit titles and snippets and produces a short answer. That answer is scored against the gold cell.
The question does not name the gold amount, person, acquirer, product, or address.
Dataset contents
100 examples falling in one of these categories:
hires / appointments financing amounts product launches acquirers new offices headcount cuts
Each row has two columns:
question— the single query sent to the search API.answer— the gold cell (amount, person, product, acquirer, location, or headcount).
This Hub dump does not include the source URL to avoid vendor overffiting indexing on the sources.
Dataset design note
Questions of the form "what product did X launch" or "how much did they raise in the latest round" are ambiguous when the same week has two real answers. Those rows were rewritten to name the round, the offering type, or the calendar day, without putting the gold string in the question.
Keep-gate: if the question already contains the gold token, the sample does not belong in the set.