TorchLLM/hindi-personaplex-7b-lora

Hindi PersonaPlex-7B LoRA Adapter

This repository contains a LoRA (Low-Rank Adaptation) adapter for the nvidia/personaplex-7b-v1 audio foundation model, specifically fine-tuned for Hindi Speech Recognition (ASR).

Performance Summary

• Final Training Loss: 0.1800
• Word Error Rate (WER): 14.2%
• Character Error Rate (CER): 5.8%
• Training Baseline: Converged from an initial loss of 0.90.

Model Description

The model was adapted using PEFT (LoRA) on a local NVIDIA Tesla V100 GPU. It leverages the 7B parameter foundation of PersonaPlex to handle Hindi linguistic nuances while remaining highly efficient (~497KB adapter size).

Usage

You can load this adapter using peft and transformers:

from transformers import AutoModel
from peft import PeftModel
import torch

base_model_id = "nvidia/personaplex-7b-v1"
adapter_id = "TorchLLM/hindi-personaplex-7b-lora"

# Load base model
model = AutoModel.from_pretrained(base_model_id, torch_dtype=torch.float16, device_map="auto")

# Load adapter
model = PeftModel.from_pretrained(model, adapter_id)

Hyperparameters

• LoRA Rank (r): 4
• LoRA Alpha: 8
• Learning Rate: 1e-4
• Precision: FP16