Queries a benchmark's image matchups — a page of compared image pairs with how the responses split between them.
An item is one image pair as one leaderboard saw it — a pair compared on several leaderboards comes back once per board. Filters use the prompt-ratings query's field set, applied to the votes a matchup is tallied from, plus four outcome fields accepted only as top-level and leaves: WinnerParticipantId/LooserParticipantId pick the side that won the whole matchup (a tie matches neither), and Responses/Extremeness select on the evidence the matchup carries — together a fair-coin sign test, so extremeness[gte]=1.96 with responses[gte]=5 keeps splits a coin flip would reproduce under 5% of the time. WinnerSigma is sigma signed from the winning side; sort by -winner_sigma to separate records the pooled data confirms from upsets. Matchups of leaderboards excluded from the benchmark overall are never returned.
Path parameters
The id of the benchmark.
Query parameters
Whether to report each side's wins weighted by annotator reliability.
What each rated entry of a prompt represents.
What a compared image stands for; per participant, self-matchups leave the page.
The 1-based page index.
The number of items per page.
Sort fields. Prefix with - for descending order (e.g. -created_at).
Filter by country.
Filter by language.
Filter by gender.
Filter by age_bucket.
Filter by occupation.
Filter by tags.
Filter by participant_id.
Filter by leaderboard_id.
Filter by run_id.
Filter by prompt_identifier.
Filter by voted_at.
Filter by winner_participant_id.
Filter by looser_participant_id.
Filter by responses.
Filter by extremeness.
Filter by winner_sigma.
How to combine the field filters: "and" (default) requires every filter to match, "or" requires any of them to match.
Response
OK