---
license: apache-2.0
---

---

Data source:
https://huggingface.co/datasets/gpt-omni/VoiceAssistant-400K

1. Question and answer audios are extracted, which results in `940,108` audio files.
2. Raw conversations are formed as multi-round chat format in `data.jsonl`, which has in total of `251,223` samples.
```
[
    ...
    {
        "messages": [
            {
                "role": "user",
                "content": "...<|audio|>"
            },
            {
                "role": "assistant",
                "content": "...<|audio|>"
            }
            {
                "role": "user",
                "content": "...<|audio|>"
            },
            {
                "role": "assistant",
                "content": "...<|audio|>"
            },
            ...
        ],
        "audios": ["path/to/first/audio", "path/to/second/audio", "path/to/third/audio", "path/to/forth/audio", ...],
    },
    ...
]
```

---