Title

Domain-Specific Fine-Tuning of PersonaPlex-7b for Customer Persona Simulation

Description

We fine-tuned NVIDIA's PersonaPlex-7b-v1 model on 200 synthetic coffee shop customer conversations to address hallucinations and poor task adherence in customer-facing deployments. Using LoRA training with ChatterboxTTS audio and LibriSpeech voices, we introduced a semantic-weighted loss function and voice prompt injection mechanism to improve emotional accuracy and reduce role-inversion hallucinations. LLM-as-judge evaluation shows improvements over the base model across all three tested configurations.

Resource type

Preprint

Resource language

English

Date created

February 2026

Date modified

February 2026

Contributors