🤖 AI Model Trainer v0.9.1

A complete GUI application for training transformer language models and exporting them to GGUF or ONNX format — from training to interactive chat testing, all in one place.

✨ What Makes This App Special

This tool provides an end-to-end workflow: select a model, train it on your dataset, export to your preferred format (GGUF or ONNX), and immediately test it in an interactive chat — all without leaving the application.

📦 Installation

Prerequisites

• Python 3.8+ (3.10+ recommended)
• Git (for cloning llama.cpp)

Step 1: Clone the Repository

git clone https://github.com/avsDeveloper/ONNX-Model-Trainer.git
cd ONNX-Model-Trainer

Step 2: Install Python Dependencies

pip install torch transformers datasets numpy psutil accelerate
pip install onnx onnxruntime optimum
pip install onnxruntime-genai  # For ONNX model inference

For GPU support (NVIDIA):

pip install torch --index-url https://download.pytorch.org/whl/cu118
pip install onnxruntime-gpu

Step 3: Set Up llama.cpp (Required for GGUF)

# Clone llama.cpp into the project directory
git clone https://github.com/ggerganov/llama.cpp.git

# Build llama.cpp
cd llama.cpp
make -j$(nproc)  # Linux/macOS
# or: cmake -B build && cmake --build build --config Release  # Windows
cd ..

Step 4: Run the Application

python trainer.py

🖥️ Platform-Specific Notes

Linux

sudo apt install python3-tk  # If tkinter is missing

macOS

brew install python-tk

Windows

• Python from python.org includes tkinter by default
• For llama.cpp, use Visual Studio or MinGW to build

📖 Understanding Model Formats

GGUF (GPT-Generated Unified Format)

• What it is: A binary format designed for efficient CPU and GPU inference with llama.cpp
• Best for: Local deployment, edge devices, CPU inference, privacy-focused applications
• Pros: Small file sizes with quantization, fast inference, no Python required for deployment
• Quantization options: F16 (full precision), Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q3_K_M, Q2_K (smallest)

ONNX (Open Neural Network Exchange)

• What it is: An open format for representing machine learning models
• Best for: Cross-platform deployment, cloud services, ONNX Runtime integration
• Pros: Wide ecosystem support, hardware acceleration, framework interoperability
• Quantization options: QInt8, QUInt8 with Dynamic quantization

⚙️ Application Parameters

Training Parameters

Parameter	Description	Default
Epochs	Number of complete passes through the dataset	3
Batch Size	Samples processed before updating weights	4
Learning Rate	Step size for weight updates	5e-5
Max Length	Maximum token sequence length	128
Save Steps	Checkpoint save frequency	500
Warmup Steps	Gradual learning rate increase steps	100
Scheduler	Learning rate schedule (linear, cosine, constant)	linear
Gradient Norm	Maximum gradient magnitude for clipping	1.0
Weight Decay	L2 regularization strength	0.01

Training Presets

• Quick Test: Fast iteration for testing (1 epoch, small batch)
• Balanced: Good balance of speed and quality
• Quality Focus: Better results, longer training
• Memory Saver: For limited VRAM/RAM systems
• Large Dataset: Optimized for big datasets
• Fine-tune: Gentle training for pre-trained models

Export Parameters (GGUF)

Parameter	Description
Quantization Type	Compression level (F16 to Q2_K)
Auto-fix EOS	Automatically fix end-of-sequence token for chat models

Export Parameters (ONNX)

Parameter	Description
ONNX Opset	ONNX operation set version (11-17)
Quant Format	Quantization format (QInt8/QUInt8)
Quant Method	Quantization method (Dynamic)
Per-Channel	Better quality quantization
Reduce Range	Better hardware compatibility

Generation Parameters (Testing)

Parameter	Description	Range
Max Tokens	Maximum response length	1-2048
Temperature	Randomness (lower = focused)	0.0-2.0
Top-P	Nucleus sampling threshold	0.0-1.0
Top-K	Number of top tokens to consider	1-100
Repetition Penalty	Penalize repeated tokens	1.0-2.0

🔄 Workflow

1. Training Tab

1. Select Base Model: Choose from supported models (GPT-2, DialoGPT, Qwen, Phi, etc.)
2. Choose Actions:
   • ☑️ Train — Fine-tune on your dataset
   • ☑️ Export — Convert to GGUF/ONNX
   • ☑️ Quantize — Compress the model
3. Configure Training: Select preset or customize parameters
4. Browse Dataset: Select your JSON dataset file
5. Start Training: Click "Start Training" or "Convert & Export"

2. Testing Tab

1. Select Format: Choose GGUF or ONNX
2. Quick Select: Pick from your exported models
3. Choose Mode: Chat, Q&A, Text Generation, etc.
4. Chat: Enter prompts and interact with your model

📁 Dataset Format

Create a JSON file with conversation pairs:

[
  {"input": "Hello!", "output": "Hi there! How can I help you?"},
  {"input": "What's the weather?", "output": "I don't have weather data, but I can help with other questions."}
]

📂 Output Structure

output/
└── ModelName/
    ├── 1_trained/           # Trained model files
    │   ├── model.safetensors
    │   ├── config.json
    │   └── tokenizer.json
    └── 2_converted/         # Exported models
        ├── model.gguf       # GGUF format
        └── model.onnx       # ONNX format

🛠️ Supported Models

Small Chat Models (Recommended)

Model	Size	GGUF	ONNX	Notes
SmolLM 135M/360M/1.7B Instruct	135M-1.7B	✅	✅	HuggingFace compact chat models
SmolLM2 135M/360M/1.7B Instruct	135M-1.7B	✅	✅	Improved 2nd generation
Qwen2 0.5B/1.5B	0.5B-1.5B	✅	✅	Alibaba's efficient models
Qwen2.5 0.5B/1.5B/3B Instruct	0.5B-3B	✅	✅	Latest Qwen, excellent quality
TinyLlama 1.1B Chat	1.1B	✅	✅	Compact Llama-based chat
StableLM 2 1.6B / Zephyr 1.6B	1.6B	✅	✅	Stability AI chat models
StableLM Zephyr 3B	3B	✅	✅	Larger StableLM variant
Phi-1/1.5/2	1.3B-2.7B	✅	✅	Microsoft code-focused
Phi-3 Mini 4K Instruct	3.8B	✅	✅	Microsoft's latest compact
MiniCPM 2B	2B	✅	❌	OpenBMB compact chat

Base/Legacy Models (Training & ONNX only)

Model	Size	GGUF	ONNX	Notes
GPT-2 / DistilGPT-2	82M-1.5B	❌	✅	Great for learning
DialoGPT Small/Medium/Large	117M-774M	❌	✅	Conversational
GPT-Neo 125M/1.3B	125M-1.3B	❌	✅	Open source GPT
OPT 125M/350M	125M-350M	❌	✅	Meta's open models

Larger Models

Model	Size	GGUF	ONNX	Notes
Gemma 2B/7B	2B-7B	✅	✅	Google's open models
Llama 2 7B	7B	✅	⚠️	Requires authentication
Mistral 7B	7B	✅	⚠️	High quality 7B

❓ Troubleshooting

"llama.cpp not found"

Build llama.cpp in the project directory:

cd llama.cpp && make -j$(nproc)

"CUDA out of memory"

• Reduce batch size
• Use "Memory Saver" preset
• Enable gradient checkpointing

"Model generates garbage"

• Ensure proper chat template is applied
• Check tokenizer files are preserved
• Try different generation parameters

📄 License

MIT License — Use at your own risk.

---

This app was primarily written with AI assistance, intended to train AI models using AI-generated datasets. These models will be used by AI-based apps. 🤖