UNO-Bench is the first unified full-modal evaluation benchmark released by Meituan's LongCat team in 2025. The related paper is titled "UNO-Bench: A Unified Benchmark for Exploring the Compositional Law Between Uni-modal and Omni-modal in Omni ModelsThe goal is to efficiently assess single-modal and multi-modal understanding capabilities.

This dataset contains 1250 full-modal samples with 98% cross-modal solvability and 2480 single-modal samples, covering 44 task types and 5 modality combinations. The dataset also includes a general scoring model that supports automated evaluation of 6 question types, providing a unified evaluation standard for multimodal tasks. The full-modal samples were carefully constructed by humans to closely resemble real-world applications, especially suitable for the Chinese context; the single-modal samples supplement the basic cognitive and ability dimensions, making the overall evaluation more comprehensive.

Data Structures:

The data is stored in Parquet format, and each sample contains structured fields:

• qid (sample ID), subset_name (subset name);
• question (textual question) and answer (standard answer);
• images / audios / videos (multimodal content, file paths are stored as a dictionary, null if not present);
• task (44 task tags), ability (ability type), source (data source), score_type (scoring method).