Empowering Long-form Omni-modal Understanding with Robust Audio Perception

We propose **AVDC (Audio-Visual Decoupled Captions)** dataset, a large-scale dataset designed to disentangle visual and auditory semantics, to improve omni understanding in multimodal models. While recent multimodal models have achieved strong performance in vision-language tasks, they often struggle with fine-grained audio-visual alignment, mainly due to the lack of such structured data. This project addresses that gap by providing both data and training pipelines for better omni-modal perception. ## Dataset ### Pipeline We propose an automatic pipeline for audio-visual decoupled caption generation, where multiple audio and language models extract and verify auditory information, temporally align it with visual content to produce segment-level captions, and finally aggregate them into a coherent global caption with subsequent verification and refinement.

### Example


Example of a multiple-choice problem. Time-related content is shown in red, and visual and audio cues in orange and blue.

## How to train 1. Prepare the dataset following `data/example.json`. 2. Download LLaVA-OneVision Model from [huggingface](https://huggingface.co/Qwen/Qwen2.5-Omni-7B). 3. Modify the parameters in `sft_omni_eval.sh`. 4. Run `bash sft_omni_eval.sh`. ## Citation If you find WorldSense helpful for your research, please consider citing our work. Thanks! ```bibtex ```