switlm 0.1.0

A simple yet powerful library to create and train custom LLMs with GGUF export

SwitLM - Simple Witty Language Model Training

SwitLM is a simple yet powerful Python library for creating and training custom Language Models with minimal code. Train your own LLM and export to GGUF format in just a few lines!

✨ Features

• 🚀 Simple API: Create and train LLMs with 3 lines of code
• 🎯 Multiple Model Sizes: From 50M to 3B parameters
• 🏗️ Modern Architecture: RoPE, RMSNorm, SwiGLU (LLaMA-style)
• 📊 Rich Dataset Support: WikiText, AG News, IMDB, SQuAD, TinyStories, and more
• 💾 GGUF Export: Direct export to GGUF format for llama.cpp
• ⚡ GPU Optimized: Memory-efficient training with gradient checkpointing
• 📈 W&B Integration: Built-in experiment tracking
• 🎨 Flexible Configuration: Easy customization of architecture and training

📦 Installation

pip install switlm

Or install from source:

git clone https://github.com/yourusername/switlm.git
cd switlm
pip install -e .

Requirements

• Python >= 3.8
• PyTorch >= 2.0.0
• transformers >= 4.30.0
• datasets >= 2.12.0

🚀 Quick Start

Basic Usage

from switlm import SwitLMTrainer

# Create and train a 1B parameter model
trainer = SwitLMTrainer(
    n_parameters="1B",
    dataset="wikitext",
    num_layers=24
)

# Train the model
trainer.train()

# Save as GGUF (ready for llama.cpp)
trainer.save("my_model.gguf")

Generate Text

# Generate text with your trained model
output = trainer.generate(
    "The future of artificial intelligence",
    max_length=100,
    temperature=0.7
)
print(output)

Multiple Datasets

# Train on multiple datasets sequentially
trainer = SwitLMTrainer(
    n_parameters="500M",
    dataset=["wikitext", "ag_news", "imdb"],
    num_layers=16
)
trainer.train()
trainer.save("multi_dataset_model.gguf")

Custom Configuration

from switlm import SwitLMTrainer, ModelConfig, TrainingConfig

# Custom model architecture
model_config = ModelConfig(
    n_parameters="custom",
    num_layers=20,
    hidden_size=1536,
    num_heads=12,
    intermediate_size=6144
)

# Custom training settings
training_config = TrainingConfig(
    learning_rate=1e-4,
    batch_size=4,
    num_epochs=3,
    use_wandb=True
)

trainer = SwitLMTrainer(
    model_config=model_config,
    training_config=training_config,
    dataset="wikitext"
)

trainer.train()
trainer.save("custom_model.gguf")

🎯 Available Model Sizes

Size	Parameters	Layers	Hidden Size	Heads	Use Case
50M	~50M	8	512	8	Quick experiments
100M	~100M	12	768	12	Small projects
500M	~500M	16	1024	16	Medium tasks
1B	~1B	24	2048	16	Serious applications
3B	~3B	32	2560	32	Production use

📚 Supported Datasets

• wikitext - Wikipedia articles
• ag_news - News classification
• imdb - Movie reviews
• squad - Question answering
• tiny_stories - Short stories
• openwebtext - Web text
• c4 - Colossal Clean Crawled Corpus
• bookcorpus - Books
• pile - Diverse text corpus

🔧 Advanced Features

Load and Continue Training

# Load a previously trained model
trainer = SwitLMTrainer(n_parameters="1B")
trainer.load("my_model.pt")

# Continue training on new data
trainer.datasets = ["squad", "imdb"]
trainer.train(num_epochs=2)
trainer.save("continued_model.gguf")

Custom Text Generation

from switlm import TextGenerator

# Create generator
generator = TextGenerator(trainer.model, trainer.tokenizer)

# Generate with custom parameters
text = generator.generate(
    "Once upon a time",
    max_length=200,
    temperature=0.8,
    top_p=0.95,
    top_k=50,
    repetition_penalty=1.2
)
print(text)

Save in Multiple Formats

# Save as both PyTorch and GGUF
trainer.save("my_model", format="both")

# Outputs:
# - my_model.pt (PyTorch checkpoint)
# - my_model.gguf (GGUF format)

🏗️ Architecture Details

SwitLM implements a modern transformer architecture with:

• RoPE (Rotary Position Embeddings): Better positional encoding
• RMSNorm: More stable training than LayerNorm
• SwiGLU: Advanced activation function (from PaLM/LLaMA)
• Pre-norm Architecture: Better gradient flow
• Gradient Checkpointing: Memory-efficient training
• Mixed Precision: Faster training with FP16

📊 Training Configuration

TrainingConfig(
    learning_rate=3e-4,         # Learning rate
    weight_decay=0.01,          # Weight decay for regularization
    beta1=0.9,                  # Adam beta1
    beta2=0.95,                 # Adam beta2
    warmup_ratio=0.1,           # Warmup ratio
    max_grad_norm=1.0,          # Gradient clipping
    batch_size=4,               # Batch size (auto if None)
    gradient_accumulation_steps=8,  # Accumulation steps
    max_length=512,             # Max sequence length
    num_epochs=1,               # Number of epochs
    use_wandb=True,             # W&B logging
    wandb_project="switlm"      # W&B project name
)

💡 Tips and Best Practices

1. Start Small: Begin with 50M or 100M models to test your pipeline
2. GPU Memory: Larger models require more VRAM (1B model needs ~12GB)
3. Dataset Size: More data generally means better models
4. Learning Rate: Start with 3e-4, adjust based on loss curves
5. Sequence Length: Shorter sequences (256-512) train faster
6. Gradient Accumulation: Increase if you run out of memory

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Inspired by modern LLM architectures (LLaMA, GPT, PaLM)
• Built with PyTorch and HuggingFace Transformers
• GGUF format support for llama.cpp integration

🌟 Star History

If you find SwitLM useful, please consider giving it a star! ⭐

---

Made with ❤️ by the SwitLM team (Avijit Paul)