[Feature Request] PersonaPlex-7B: Fine-Tuning, RAG & Custom Voice Support
Hi 👋 I’m using nvidia/personaplex-7b-v1 to build a real-time voice customer support agent and have a few questions:
- Fine-tuning
Are fine-tuning scripts/docs planned?
Supported data formats (audio + transcripts)?
Hardware requirements?
LoRA / QLoRA support?
- Custom knowledge integration
Max context/token limit for prompts?
Can context be updated dynamically during a session?
Recommended RAG setup (ASR → RAG → PersonaPlex)?
Latency considerations for real-time/full-duplex use?
Is domain fine-tuning supported/planned?
- Custom voice
Custom voice embeddings or voice cloning?
Required audio format and duration?
Current workaround: static prompt-based knowledge
Limitations: token limits, no retrieval, manual updates
Env: A10G 24GB | production voice agent
Thanks! Any guidance or roadmap info would be appreciated.
Hey I have been trying to use the model for some use-cases and want to know is there any official doc or soemthing regarding the fine-tuning of the model.
I don't recommend using this model checkpoint for production. Its more of a showcase for naturalness. Stay tuned for future models that are smarter and are packaged with finetuning flows and toolcalling support. Can't promise custom voice though since its extremely difficult to get legal approval for that.
Any updates on this topic?
bump
:-(