Finance Search
The validation set of the OpenBenchmarks finance search benchmark, which measures how well a search API helps an agent answer a finance professional's questions. The agent, prompt, search budget and grader are held fixed; only the search API changes. Use it to develop and tune against; results on it are not leaderboard scores.
Configurations
| Configuration | Rows | Task |
|---|---|---|
historical-lookup |
40 | One published figure for a named entity and period; one well-aimed search should find it. |
multi-hop-lookup |
20 | Several published facts combined: a peer comparison, a trend, a result against guidance. |
live-lookup |
27 | A price at a given moment, asked shortly after it: 14 stocks and 13 coins. |
Each configuration has one split named val.
Fields
historical-lookup and multi-hop-lookup:
id: question id.question: the question as asked, with named entities and absolute dates.answer: the reference answer. Where it ends with "Wrong:", the listed values are the expected wrong answers.tolerance: how close a numeric answer must be.area: one of company results, banks and insurers, macro and policy, small and international, markets.
live-lookup:
id,asset(equity,crypto),name,ticker.query_template: the query a search API receives once filled in:<name> (<ticker>) closing price {{date}}for stocks (that day's official close) and<name> (<ticker>) price {{date}} {{time}}for coins (time in UTC).
Live lookup
Live lookup measures freshness: how soon a new price reaches a search API's results. In the benchmark each item is asked shortly after a moment that has just passed and graded against market data pulled afterwards, so a live answer only exists once that moment has passed. These rows carry the items and the query format, without answers.
Grading
An answer is correct only if every part the question asks for is right within its tolerance. A hedged answer counts only if its committed value is right.
License
No license is assigned yet.
- Downloads last month
- 31