Private dataset generated from Twentle Twenty Questions self-play logs.
train: 11126 examples
validation: 2794 examples
Source games: 499
Validation target fraction: 0.2
Split seed: 42
Split policy: game-level split to avoid train/validation leakage
Questioner first turn: skipped
Example format: system, user, assistant
data/twenty_questions_games.jsonl
