OmniBench
Multimodal · 2024-09-23
OmniBench tests tri-modal understanding with 1,142 human-annotated four-option questions, each pairing an image with an audio clip (771 speech, 265 sound-event and 106 music items). Scores are accuracy over all 1,142 questions using the official prompt and answer parser.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-Omni-2B | 46.3 |
| Phi-4-multimodal-instruct | 39.0 |