YANAMI: A Large-Scale Multimodal Dataset and Benchmark for Multiturn Stateful Interaction
Fine-tuning BAGEL on YANAMI improved multimodal stateful interaction by 13.23%, the researchers report.
YANAMI contains 400,000 multi-turn dialogues, more than 2 million images and 2.6 million dialogue turns across image fusion, editing, world modeling and visual reasoning. Its tasks require models to preserve and revise information across turns. A separate 400-dialogue benchmark tests both understanding of the evolving dialogue state and faithful rendering in visual outputs. Existing unified multimodal models still struggle with both, according to the paper.
HF Daily Papers' note
YANAMI contains 400,000 multi-turn dialogues, more than 2 million images and 2.6 million dialogue turns across image fusion, editing, world modeling and visual reasoning. Its tasks require models to preserve and revise information across turns. A separate 400-dialogue benchmark tests both understanding of the evolving dialogue state and faithful rendering in visual outputs. Existing unified multimodal models still struggle with both, according to the paper.
HF Daily Papers' note
score 5