Skip to content

USER_GUIDE.md

Latest commit

 

History

History
171 lines (119 loc) · 3.9 KB

USER_GUIDE.md

File metadata and controls

171 lines (119 loc) · 3.9 KB
Raw

SLMGEN User Guide

Generate fine-tuning notebooks for Small Language Models in minutes.

Quick Start

1. Sign Up / Login

Visit localhost:3000/signup to create an account or login with GitHub.

2. Upload Dataset

  • Click Get Started on the dashboard
  • Drag and drop your JSONL file or click to browse (supports live client-side chat preview for first 10KB)
  • Required format: ChatML with messages array
{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

3. Review Analysis

After upload, you'll see:

  • Total Examples - Number of training examples
  • Total Tokens - Approximate token count
  • Quality Score - Dataset quality rating (0-100%)

4. Select Task Type

Choose your fine-tuning goal:

  • Instruction Following - General task completion
  • Chat/Conversation - Multi-turn dialogue
  • Code Generation - Programming tasks
  • Summarization - Text condensation
  • Classification - Category assignment

5. Choose Deployment Target

Where will your model run?

  • Cloud (A100/H100) - Maximum performance
  • Cloud (T4/L4) - Budget-friendly GPU
  • Edge/Mobile - Optimized for devices
  • CPU Only - No GPU required

6. Get Recommendations

SLMGEN analyzes your dataset and requirements to recommend the best models:

  • Score - Overall fit percentage
  • VRAM - GPU memory required
  • Training Time - Estimated duration

7. Configure Training

Select a preset:

  • Fast (~15 min) - Quick experimentation
  • Balanced (~45 min) - Best for most cases
  • Quality (~2 hours) - Maximum quality

Or customize:

  • LoRA Rank, Learning Rate, Epochs, Batch Size, etc.

8. Generate Notebook

Click Generate Notebook to create your Colab notebook.

During generation, you'll see a Training Terminal Simulator that visualizes the Unsloth training commands in real-time. Once complete, a Success Confetti celebration confirms your notebook is ready.

Options:

  • Open in Colab - Start training immediately
  • Download - Save locally

Dataset Format

ChatML (Recommended)

{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is 2+2?"},
    {"role": "assistant", "content": "2+2 equals 4."}
  ]
}

Tips for Quality Data

  1. Minimum 100 examples for basic fine-tuning
  2. 1,000+ examples recommended for best results
  3. Consistent formatting across all examples
  4. Include system prompts for better control
  5. Mix easy and hard examples for robustness

Model Selection Guide

Model Size Best For VRAM
Phi-4 Mini 3.8B Instruction, Reasoning 8GB
Llama 3.2 1-3B General, Chat 4-8GB
Gemma 2 2B Efficient fine-tuning 4GB
Qwen 2.5 3B Multilingual 8GB
DeepSeek Coder 1.3B Code generation 4GB

Training Presets

Fast

  • LoRA Rank: 8
  • Learning Rate: 5e-4
  • Epochs: 1
  • Time: ~15 minutes

Balanced

  • LoRA Rank: 16
  • Learning Rate: 2e-4
  • Epochs: 3
  • Time: ~45 minutes

Quality

  • LoRA Rank: 32
  • Learning Rate: 1e-4
  • Epochs: 5
  • Time: ~2 hours

Export Options

After training, export your model to:

  • HuggingFace Hub - Share publicly or privately
  • ONNX - For edge deployment
  • Ollama - Run locally with Ollama
  • GGUF - Quantized for llama.cpp

Troubleshooting

"Dataset quality score is low"

  • Check for duplicate examples
  • Ensure consistent message formatting
  • Add more diverse examples

"Out of memory during training"

  • Reduce batch size
  • Lower LoRA rank
  • Use a smaller model
  • Enable gradient checkpointing

"Model not converging"

  • Increase epochs
  • Reduce learning rate
  • Check data quality

Support