OmniVideo-100K, a 100K-sample multimodal dataset featuring paired audio and video with structured reasoning scripts and evidence chains, is now available on HuggingFace. The dataset targets audio-visual reasoning tasks where models must align temporal events across modalities. For multimodal model developers, this dataset reduces friction in training video understanding systems beyond vision-only architectures. Audio-visual alignment remains underexplored relative to image-text work, making structured reasoning chains particularly valuable for supervising reasoning steps rather than only final answers. Teams previously cobbling together proprietary datasets or smaller public sources can now standardize on a common benchmark. Operationally, this lowers the cost of entry for audio-visual reasoning research and enables clearer cross-team comparisons. Models trained on OmniVideo-100K create a reference point for measuring progress on joint modality understanding—critical for applications requiring temporal coherence across sound and video (surveillance, embodied AI, video QA systems). Expect downstream fine-tuning on this dataset to become routine in video-centric pipelines where audio currently remains underutilized.