README.md

metadata

title: Nemotron LoRA Fine-tuning
emoji: 🚀
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 4.0.0
app_file: app.py
pinned: false
license: apache-2.0

🚀 LoRA Fine-tuning for Nemotron-3-8B

Train Nvidia's Nemotron-3-8B model with LoRA on your custom datasets.

Features

• ✅ 4-bit Quantization with QLoRA
• ✅ PEFT Integration
• ✅ Gradio UI
• ✅ Auto Push to Hub
• ✅ Progress Tracking

Quick Start

1. Upgrade to GPU (Settings → Hardware)
2. Configure training parameters
3. Add HF token if needed
4. Click "Start Training"

Configuration

Default settings are optimized for A10G (24GB VRAM). Adjust for your GPU:

• T4 (16GB): batch_size=2, gradient_accumulation=8
• A10G (24GB): batch_size=4, gradient_accumulation=4
• A100 (40GB): batch_size=8, gradient_accumulation=2

Dataset Format

Supports multiple formats (auto-detected):

• Q&A: {"question": "...", "answer": "..."}
• Instruction: {"instruction": "...", "response": "..."}
• Text: {"text": "..."}

Using the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("nvidia/nemotron-3-8b-base-4k")
model = PeftModel.from_pretrained(base_model, "YOUR_USERNAME/nemotron-lora")
tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/nemotron-lora")

Support

For issues: GitHub