Triangulum 5 B

Introduction

Triangulum-5B is a multilingual collection of pretrained and instruction-tuned generative models designed for text generation tasks across multiple languages. These models are optimized for complex reasoning tasks using an autoregressive transformer architecture.

Architecture

Triangulum-5B builds upon LLaMA's autoregressive language model, using an optimized transformer framework. This foundation enhances performance, supports multiple languages, and is tuned via supervised and reinforcement learning methods to align with human preferences.

Training

1. Synthetic Datasets: Uses chain-of-thought datasets to enhance reasoning capabilities.
2. Supervised Fine-Tuning (SFT): Aligns the model with specific tasks through curated data.
3. Reinforcement Learning with Human Feedback (RLHF): Iteratively trains the model to adhere to human values and safety guidelines.

Guide: Running Locally

Step 1: Download the Model

Use Ollama to download the Triangulum-5B model:

ollama run triangulum-5b-f16.gguf

Step 2: Model Initialization and Download

Ollama will initialize and download the model files. Monitor the terminal for progress and confirmation.

Step 3: Interact with the Model

Once ready, interact with the model by sending prompts, such as:

>>> What can you do for me?

Step 4: Exit the Program

To exit, type:

/exit

Cloud GPUs: For optimal performance, consider using cloud GPUs to handle the computational demands of running large models like Triangulum-5B locally.

License

Triangulum-5B is available under the CreativeML OpenRAIL-M license, permitting use in compliance with specified guidelines.