---
license: other
pretty_name: DuplexOmni Data
size_categories:
- 1B<n<10B
task_categories:
- any-to-any
- automatic-speech-recognition
- text-to-speech
- visual-question-answering
---

# DuplexOmni Data

This dataset accompanies **DuplexOmni: Real-Time Listening, Seeing, Thinking, and Speaking for Full-Duplex Interaction**.

- GitHub: [MuyeHuang/DuplexOmni](https://github.com/MuyeHuang/DuplexOmni)
- arXiv: [2606.09186](https://arxiv.org/abs/2606.09186)

## Files

The root directory includes the Writer-Director source files:

- `inbound.director.jsonl`
- `outbound.director.jsonl`

The synthesized parquet training shards are about **9 TB** in total, so uploading them is expected to be very slow. They will be uploaded incrementally; each completed shard can be used as soon as it appears.

Before all parquet shards finish uploading, users can synthesize the data locally by following the method and scripts in [MuyeHuang/DuplexOmni](https://github.com/MuyeHuang/DuplexOmni).