kcz358/aero-realtime-data

Aero Realtime Data

Training data for Aero Realtime, a native proactive audio-video language model whose input and output share one aligned autoregressive stream.

⚠️ This is only part of the data

This repository hosts only the self-packaged assets that are not available on an official Hugging Face release:

• qaego4d videos and TTS audio
• all TTS audio for EgoIT-99K
• the lmms_engine_parquet / sft_parquet manifests used by the training configs

The full dataset additionally pulls from these official public releases, which are downloaded separately to keep this repo lightweight and avoid re-hosting:

• lmms-lab/LLaVA-Video-178K
• lmms-lab/EgoIT-99K
• chenjoya/Live-WhisperX-526K

Preparing the full dataset

Use the one-shot preparation script in the code repository to assemble the complete blob layout the stage1/stage2 configs expect. It downloads the sources above, merges in the assets from this repo, and extracts every archive in place.

git clone https://github.com/kcz358/aero-realtime.git
cd aero-realtime

bash scripts/data/prepare_data.sh --dest /scratch/aero-ws/data

Re-running is cheap: finished datasets are skipped via marker files. Select a subset with --only llava,egoit,livecc,assets or force a refresh with --force. See scripts/data/prepare_data.sh --help and the header of that script for the full environment overrides and the resulting on-disk layout.

Citation

If you find this work useful, please cite:

@misc{zhang2026aerorealtimefullyaligned,
      title={Aero Realtime: Fully Aligned Input-Output Streams for Low-Latency Streaming Multimodal Generation},
      author={Kaichen Zhang and Wei Huang and Keming Wu and Bo Li and Xiaojuan Qi},
      year={2026},
      eprint={2608.08469},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.08469}
}