Triangulum 1 B

Introduction

Triangulum-1B is a collection of multilingual large language models (LLMs) designed for generative text applications. These models are pretrained and instruction-tuned for complex reasoning tasks, making use of synthetic datasets based on long chains of thought.

Architecture

• Foundation Model: Built on LLaMA's autoregressive language model, optimized with a transformer architecture for improved performance.
• Instruction Tuning: Incorporates supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align outputs with human preferences for helpfulness and safety.
• Multilingual Support: Capable of handling several languages including English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Training

• Synthetic Datasets: Utilizes data based on long chains of thought to enhance reasoning capabilities.
• Supervised Fine-Tuning (SFT): Aligns the model to specific tasks using curated datasets.
• Reinforcement Learning with Human Feedback (RLHF): Ensures adherence to human values and safety through iterative training.

Guide: Running Locally

Example 1: Running Triangulum-1B

1. Download the Model: Use the command ollama run triangulum-1b-f16.gguf to download the model.
2. Model Initialization and Download: Ollama will automatically initialize and download the model files.
3. Interact with the Model: Once ready, send prompts to the model for text generation tasks.
4. Exit the Program: Type /exit to close the session.

Notes on Using Quantized Models

• Quantized models like triangulum-1b-f16.gguf are optimized for resource-constrained hardware.
• Ensure sufficient VRAM or CPU resources are available.
• Use the .gguf model format for compatibility with Ollama.

Cloud GPU Recommendation

For enhanced performance and scalability, consider using cloud GPU services like AWS, Google Cloud, or Azure.

License

Triangulum-1B is licensed under the CreativeML Open RAIL-M, ensuring open access while maintaining responsible AI use.