Open-Jev (dataset)

The training and evaluation data behind open-jev-9b and its 2B/27B siblings: 12 configs, 520,199 rows total across 60 splits (configs overlap heavily, so the totals don't sum to a clean whole). Mostly synthetic and controlled by the dataset card's own description — customer-support routing, workflow decisions, game states (Snake, tic-tac-toe), citation-support judgments, entity/amount/email/phone extraction, and an 11,600-row multilingual (English/Chinese/Turkish) mailroom config. CC0-1.0 on the generated content only; the card is explicit that this "does not cover upstream wording, external assets, source data or model weights," and that citation/entity/ extraction configs were never actually used for training or inference — held out as controls.

The release-v2 split open-jev-9b trains on (80,816 rows) is one slice of this. Its own "OOD" split is source-specific variation within the generator's own configs, not held-out real-world domains — a narrower claim than "out of distribution" usually means. Read the OOD numbers with that scope in mind wherever they get quoted.

Related

open-jev-9b · laya · zefan-cai