Adapt NVIDIA PersonaPlex-7B-v1 for Russian, Uzbek

Job ID: 40302600

Budget: $250 – $750 USD

Title

Senior ML Engineer – Fine-tuning NVIDIA PersonaPlex (Helium) for Russian and Uzbek

Project Description

We are building a real-time voice AI assistant based on NVIDIA PersonaPlex-7B-v1.

The model currently works in English.
Our goal is to adapt it for Russian and Uzbek and integrate it into a voice assistant platform.

We already have prepared datasets.

We are looking for an experienced Speech / LLM engineer who has worked with multimodal models and LoRA fine-tuning.

---

Main Tasks

1. Fine-tune NVIDIA PersonaPlex-7B-v1 for Russian and Uzbek language support
2. Train LoRA adapters for multilingual capability
3. Adapt the speech-to-speech pipeline (audio tokens + Helium backbone)
4. Integrate persona conditioning for our voice assistant
5. Optimize inference for real-time interaction

---

Model Architecture

Base model:

NVIDIA PersonaPlex-7B-v1
(Moshi architecture + Helium backbone)

Speech-to-speech conversational AI with full-duplex interaction.

---

Dataset

Prepared datasets include:

• Russian conversation dataset
• Uzbek conversation dataset
• Persona dialogue dataset

Estimated size: ~300k–500k samples.

---

Technical Requirements

Strong experience with:

• PyTorch
• HuggingFace Transformers
• LoRA / QLoRA / PEFT
• Multilingual model fine-tuning
• Speech models or audio token models

Experience with one of the following is highly preferred:

• Whisper
• NeMo
• wav2vec
• multimodal LLMs

---

Infrastructure

Training environment:

• A100 / H100 GPUsor other
• Docker environment

---

Expected Deliverables

• LoRA adapters for RU and UZ
• Training scripts
• Evaluation results
• Integration instructions

---

Project Timeline

Estimated implementation time:

3–7 days.

---

When applying please include

• examples of LLM fine-tuning projects
• experience with speech models
• experience with LoRA training