[Benchmark] Add Support for LVOmniBench

[KD-TAO]

Summary

Add LVOmniBench, a benchmark for evaluating omnimodal LLMs on long-form audio-video understanding.

• 1,014 manually constructed multiple-choice questions
• 275 long videos (10–90 minutes, avg 34 min)
• Requires joint audio-visual reasoning
• Dataset: https://huggingface.co/datasets/KD-TAO/LVOmniBench

Evaluation

• Fine-grained accuracy across 4 dimensions:
  • question_type (9 categories): Attribute Perception, Counting, Event Understanding, etc.
  • audio_type (3 categories): Speech, Music, Sound
  • difficulty (3 levels): High, Medium, Low
  • video_category (6 categories): Film & TV, Entertainment, Sports, etc.

Changes

File	Change
vlmeval/dataset/lvomnibench.py	New dataset class
vlmeval/dataset/utils/lvomnibench.py	Evaluation utilities
vlmeval/dataset/__init__.py	Registration (+2 lines)
vlmeval/dataset/video_dataset_config.py	Config variants (+11 lines)

Test

# Default (video_llm mode, no frame extraction)              
python run.py --data LVOmniBench --model Qwen2.5-Omni-7B-ForVideo