Preference
Arena 140K
Arena 140K records which of two models a person preferred after reading both replies to a prompt they wrote themselves, with ties and mutual rejection both counted as outcomes.
128,287items
53subjects
CC-BY-4.0license
preferencedomain
textmodality
item-level responses released
Saturation status: No
Response matrix
Rows are model pairs, ordered by how many items the pair was compared on; within a row the comparisons are sorted by outcome. A column position is not a shared item — arena prompts are almost never reused across pairs, so cells in the same column are different items.
Loading session strips…
lowhighUnobserved
Scale: {0, 0.5, 1} (1=win, 0.5=tie, 0=loss/both_bad)