ForgeLLM

ForgeLLM is a comprehensive platform for continued pre-training and instruction fine-tuning of large language models using MLX on Apple Silicon.

What ForgeLLM Does

• 🚀 Train: Continued pre-training (CPT) via web interface (IFT coming soon - see Development Perspectives)
• 📊 Monitor: Real-time training dashboards and checkpoint management
• 🆚 Compare: Enable comparison of multiple training sessions with validation loss, perplexity, stability and generalization gap
• 🔗 Fuse: Merge LoRA/DoRA adapters with base models for deployment
• ⚡ Quantize: Convert models to 8-bit or 4-bit precision for efficient deployment
• 💬 Chat & Test: Interactive chat with models and adapters via CLI or web
• 📦 Publish: Convert and publish trained models with comprehensive documentation

Screenshots

Training:

Monitoring:

Compare:

Testing:

Quick Start

1. Installation

Option A: Install from PyPI (Recommended)

# Install latest version
pip install forgellm

# Install specific version
pip install forgellm==0.4.7

# Upgrade existing installation
pip install --upgrade forgellm

Option B: Install from Source (Development)

git clone https://github.com/lpalbou/forgellm.git
cd forgellm
pip install -e .

Requirements: Python 3.9+ and Apple Silicon Mac (M1/M2/M3/M4). All dependencies including MLX are installed automatically.

2. Download Models

# Install HuggingFace CLI
pip install huggingface_hub

# Download a model (examples)
huggingface-cli download mlx-community/gemma-3-1b-it-bf16     # Small model
huggingface-cli download mlx-community/Qwen3-4B-bf16         # Medium model

3. Start ForgeLLM

# Start both servers (recommended)
forgellm start

# Opens web interface at http://localhost:5002
# Model server runs at http://localhost:5001

That's it! 🎉

Usage

Web Interface (Recommended)

The web interface provides everything you need:

forgellm start                    # Start both servers
# or
forgellm web --port 5002         # Web interface only
forgellm server --port 5001      # Model server only (separate terminal)

Web Interface Features:

• Training Tab: Configure and start CPT training (IFT support coming soon)
• Monitoring Tab: View training progress and dashboards
• Testing Tab: Chat with models and test different prompts

Command Line Interface

The CLI is perfect for quick model testing and interactive chat:

# Interactive chat with a model (REPL mode)
forgellm cli generate --model mlx-community/gemma-3-1b-it-bf16

# Single prompt test
forgellm cli generate --model mlx-community/gemma-3-1b-it-bf16 --prompt "Hello, how are you?"

# Get model architecture info
forgellm cli info --model mlx-community/gemma-3-1b-it-bf16

# Test with an adapter (your trained model)
forgellm cli generate --model mlx-community/Qwen3-4B-bf16 --adapter-path models/cpt/my_trained_model

REPL Mode Commands:

• Type normally to chat
• /help - Show available commands
• /q or /exit - Quit
• /stats - Show session statistics
• /system [prompt] - Set/show system prompt

Model Downloads

ForgeLLM works with MLX-compatible models from HuggingFace. All models are cached locally in ~/.cache/huggingface/hub/.

Recommended Models

Small Models (1-2B) - Good for testing:

huggingface-cli download mlx-community/gemma-3-1b-it-bf16
huggingface-cli download mlx-community/gemma-3-1b-pt-bf16

Medium Models (3-4B) - Good balance:

huggingface-cli download mlx-community/Qwen3-4B-bf16
huggingface-cli download mlx-community/gemma-3-4b-it-bf16

Large Models (7-8B) - Best quality:

huggingface-cli download mlx-community/Qwen3-8B-bf16
huggingface-cli download mlx-community/Meta-Llama-3.1-8B-Instruct-bf16

Model Types

• Base Models (-bf16, -pt-): Ideal for continued pre-training, clean slate for domain adaptation
• Instruct Models (-it-, -Instruct-): Can also be used for continued pre-training with careful data mixing
• Quantized Models (-4bit, -8bit): Smaller memory usage, slightly lower quality

Continued Pre-training: Base vs Instruct Models

Base Models (Recommended for CPT):

• ✅ No instruction-following capabilities to preserve
• ✅ Clean foundation for domain-specific knowledge
• ✅ Higher learning rates and longer training possible

Instruct Models (Advanced CPT):

• ✅ Better at learning from complex documents (recent research)
• ⚠️ Requires careful data mixing (1-5% original pretraining data)
• ⚠️ Lower learning rates to prevent catastrophic forgetting
• ⚠️ Shorter training to avoid losing instruction-following abilities

Choose base models for straightforward domain adaptation, instruct models when you need better knowledge absorption from complex documents.

📖 For detailed CPT best practices and latest research findings, see docs/cpt.md

Training Your Own Models

Continued Pre-Training (CPT) - Available Now

1. Prepare Data: Place text files in dataset/ directory
2. Start Web Interface: forgellm start
3. Training Tab: Configure model, data, and parameters
4. Monitor: Watch progress in real-time
5. Publish: Convert best checkpoints to full models

Training is currently only available through the web interface.

Instruction Fine-Tuning (IFT) - Coming Soon

IFT capabilities are currently in development. For technical details and implementation roadmap, see Development Perspectives.

Directory Structure

forgellm/
├── dataset/          # Your training data (text files)
├── models/           # Trained model outputs
│   ├── cpt/         # Continued pre-training models
│   └── ift/         # Instruction fine-tuning models (coming soon)
└── data/            # Processed training data

Commands Reference

Main Commands

forgellm start                    # Start both servers (recommended)
forgellm web [--port 5002]       # Web interface only
forgellm server [--port 5001]    # Model server only
forgellm cli <command>            # Command-line operations

CLI Commands

# Interactive chat (REPL mode)
forgellm cli generate --model <model>

# Single prompt
forgellm cli generate --model <model> --prompt "Your question"

# Model information
forgellm cli info --model <model>

# Test with adapter
forgellm cli generate --model <model> --adapter-path <path>

Requirements

• Hardware: Apple Silicon Mac (M1/M2/M3/M4)
• Memory: 16GB+ RAM recommended
• Storage: 5-20GB per model
• Python: 3.9+
• MLX: Automatically installed

Architecture

ForgeLLM uses a clean separation:

• Model Server (forgellm server): Handles model loading and inference
• Web Server (forgellm web): Provides UI and training coordination
• CLI (forgellm cli): Direct model interaction and testing

This allows you to use just the CLI for testing, or the full web interface for training.

Documentation

📚 Comprehensive Guides

• Getting Started: Complete setup and first training session
• Architecture: System design and component overview
• Data Flow: How data moves through the system
• API Reference: Complete REST API and CLI documentation
• CPT Best Practices: Advanced continued pre-training techniques
• Development Perspectives: Current capabilities and IFT roadmap

🔧 Technical Documentation

• Architecture: Multi-process design with model server separation
• Training Pipeline: Real-time monitoring with automatic checkpoint management
• Model Publishing: LoRA to full model conversion with comprehensive documentation
• Error Recovery: Robust error handling and automatic recovery mechanisms

Contributing

Contributions welcome! Please submit pull requests.

License

MIT License - see LICENSE file.

Acknowledgments

• ForgeLLM Team: Continued pre-training platform
• MLX-LM: Apple's MLX framework for LLMs
• MLX: Apple's machine learning framework