⚒️ Finetuning-Forge

Local-first LLM fine-tuning workstation — a modern UI over raw command-line scripts.
Real-time GPU telemetry, live loss curves, and VRAM pre-flight checks on consumer GPUs.

Architecture • Features • Quick Start • Tech Stack

---

---

🏗 Architecture

┌──────────────────────────────────────────────────────────────┐
│                     Finetuning-Forge                         │
│                                                              │
│  ┌─────────────────────────┐    ┌────────────────────────┐  │
│  │    Next.js Frontend      │    │   FastAPI Backend       │  │
│  │                         │    │                        │  │
│  │  ┌─────────────────┐   │    │  ┌──────────────────┐  │  │
│  │  │  Script Lab      │   │◄──►│  │  Orchestrator    │  │  │
│  │  │  (Monaco Editor) │   │    │  │  (Multi-Engine)  │  │  │
│  │  └─────────────────┘   │    │  └────────┬─────────┘  │  │
│  │                         │    │           │             │  │
│  │  ┌─────────────────┐   │    │  ┌────────▼─────────┐  │  │
│  │  │  GPU Dashboard   │◄──┼────┼──│  WebSocket       │  │  │
│  │  │  (Live Charts)   │   │    │  │  Streaming       │  │  │
│  │  └─────────────────┘   │    │  └────────┬─────────┘  │  │
│  │                         │    │           │             │  │
│  │  ┌─────────────────┐   │    │  ┌────────▼─────────┐  │  │
│  │  │  VRAM Pre-flight │   │    │  │  Training Backend│  │  │
│  │  │  Check           │   │    │  │  Unsloth/Axolotl │  │  │
│  │  └─────────────────┘   │    │  │  /Torchtune       │  │  │
│  └─────────────────────────┘    │  └──────────────────┘  │  │
│                                  └────────────────────────┘  │
└──────────────────────────────────────────────────────────────┘
                              │
                    ┌─────────▼──────────┐
                    │   Your GPU / CUDA  │
                    │   nvidia-smi       │
                    └────────────────────┘

Data Flow

User selects model + dataset
        │
        ▼
VRAM Pre-flight Check ──── Not enough VRAM? ──── Warning shown
        │ (OK)
        ▼
Training backend spawned (Unsloth / Axolotl / Torchtune)
        │
        ├── Live loss curves ──────────────────► Frontend charts
        ├── GPU telemetry (VRAM/temp/util) ────► GPU Dashboard  
        └── Training logs ─────────────────────► Script Lab

---

✨ Features

Feature	Description
🔧 Multi-Engine Orchestrator	Switch between Unsloth, Axolotl, and Torchtune without reconfiguring
📝 Script Lab	Monaco Editor with live-synced training scripts — edit and run in one place
📊 GPU Command Center	Real-time VRAM usage, temperature, utilization via nvidia-smi
📈 Live Loss Curves	Training loss streamed live over WebSocket — no polling
✅ VRAM Pre-flight Checks	Validates available GPU memory before starting — prevents OOM crashes
🌑 Tactical Obsidian UI	Dark mode with safety-orange accents — built for long training sessions

---

⚡ Quick Start

# Clone
git clone https://github.com/PeakScripter/Finetuning-Forge.git
cd Finetuning-Forge

# Frontend (Terminal 1)
npm install
npm run dev
# → http://localhost:3000

# Backend (Terminal 2)
cd backend
pip install -r requirements.txt
uvicorn main:app --reload --port 8001
# → http://localhost:8001

Install a Training Backend

# Pick one (or all):
pip install unsloth
pip install axolotl
pip install torchtune

---

🗂 Project Structure

Finetuning-Forge/
├── app/                  # Next.js pages & API routes
├── backend/              # FastAPI server (Python)
│   ├── main.py           # Entry point
│   └── orchestrator/     # Training backend abstraction
├── components/           # React UI components
├── context/              # Global state management
├── lib/orchestrator/     # Frontend-side backend abstraction
├── models/               # Local model files
└── datasets/             # Training data (JSONL format)

---

⚙️ Requirements

• Node.js 18+
• Python 3.10+
• NVIDIA GPU with CUDA (recommended)
• NVIDIA Drivers (nvidia-smi accessible)
• At least one training backend installed (Unsloth, Axolotl, or Torchtune)

---

🛠️ Tech Stack

Frontend: Next.js 16, React 19, Tailwind CSS, Framer Motion, Monaco Editor

Backend: FastAPI, WebSocket streaming, nvidia-smi integration

Training: Unsloth, Axolotl, Torchtune

Language: TypeScript + Python

---

📄 License

MIT License — see LICENSE for details.