prepforge 0.2.0

AI Exam Preparation System with TUI, GUI, and Streamlit

PrepForge

PrepForge is a modular AI-powered exam preparation system that provides multiple interfaces (TUI, GUI, and Web) for interacting with large language models. It supports both Hugging Face transformer models and GGUF-based models via llama.cpp, along with optional LoRA adapters and a built-in training pipeline.

---

Overview

PrepForge is designed as a unified interface for:

• Running local language models
• Managing multiple model backends (Hugging Face and GGUF)
• Applying LoRA adapters dynamically
• Generating study material (notes, MCQs, plans, practice)
• Training custom models using QLoRA / Unsloth

The system separates concerns between configuration, model loading, and interface layers.

---

Tech Stack

Category	Technologies / Tools	Purpose
Core	Python 3.10+, Transformers, PEFT, HF Hub	Model loading, LoRA support, model management
Inference Backend	PyTorch, llama-cpp-python	Running HF models (GPU/CPU) and GGUF models
Interfaces	Rich, prompt_toolkit, PyQt6, Streamlit	TUI, CLI input, GUI, and Web interface
Training	Unsloth, TRL, Datasets	Fine-tuning models using QLoRA pipelines
System Utilities	fzf (optional), subprocess, JSON	Model selection, CLI execution, config storage

---

Installation

Base Installation

pip install prepforge

Optional Features

pip install prepforge[gguf]        # GGUF (llama.cpp) support
pip install prepforge[gpu]         # CUDA / torch support
pip install prepforge[web]         # Streamlit interface
pip install prepforge[gui]         # PyQt6 GUI
pip install prepforge[train]       # Training dependencies
pip install prepforge[gguf,gpu,web,gui]

---

Architecture

CLI Layer (cli.py)

Handles:

• Running interfaces
• Model installation
• Configuration
• Training commands

Core Layer (core/)

• model.py: Loads HF and GGUF models, applies LoRA
• utils.py: Prompt building and streaming
• train.py: Fine-tuning pipeline

Interface Layer

• tui_app.py: Terminal UI (Rich)
• gui_app.py: Desktop GUI (PyQt6)
• streamlit_app.py: Web UI

Configuration Layer

Stored at:

~/.prepforce/config.json

---

Model Support

Hugging Face Models

• Loaded via transformers
• Cached automatically
• Supports LoRA

Examples:

meta-llama/Llama-3-8B
mistralai/Mistral-7B-Instruct

---

GGUF Models

• Single-file models
• Run via llama.cpp
• Lower resource usage

Example:

Qwen3-4B-Q4_1.gguf

---

LoRA Adapters

• Lightweight fine-tuned layers
• Applied on top of base HF models
• Must match base architecture

---

Model Installation

prepforge install_model

Features:

• Browse models from Hugging Face
• Filter by type (GGUF / HF / LoRA)
• Install:
  • GGUF → single file
  • HF → full snapshot
  • LoRA → adapter files

Storage:

• GGUF → ~/models
• HF → ~/.cache/huggingface/hub

---

Configuration

Set default model:

prepforge config --model <model_path_or_repo>

Set LoRA:

prepforge config --lora <adapter_path>

Interactive selection:

prepforge config

---

Running the Application

Modes:

• Chat
• Notes
• MCQ generation
• Study plan
• Practice

Terminal UI

prepforge run tui

---

Desktop GUI

prepforge run gui

---

Web Interface

prepforge run streamlit

Open:

http://localhost:8501

---

Inference Flow

1. Load model from config
2. Detect backend:
   • GGUF → llama.cpp
   • HF → transformers
3. Apply LoRA if provided
4. Build prompt
5. Generate response (streaming)

---

Training

prepforge train \
  --dataset dataset.jsonl \
  --epochs 3 \
  --output trained_model

Options:

• --limit → limit samples
• --subset → percentage of dataset
• --lr → learning rate

---

Dataset Format

{
  "instruction": "Explain Newton's laws",
  "input": "",
  "output": "Newton's laws describe motion..."
}

---

Project Structure

major_project/
├── core/
│   ├── model.py
│   ├── train.py
│   └── utils.py
├── tui_app.py
├── gui_app.py
├── streamlit_app.py
├── cli.py
├── config.py

---

Requirements

• Python 3.10+
• CUDA GPU recommended for HF models

---

Notes

• Models are not bundled
• Must be installed separately
• GGUF requires llama-cpp-python
• LoRA must match base model

---

Author

Himanshu Arora

---

License

MIT License