Llama 3.1 Tulu 3 8 B

Introduction

LLAMA-3.1-TULU-3-8B is a state-of-the-art instruction-following model developed by AllenAI. It is part of the Tülu3 model family, designed for a wide range of tasks including text generation and comprehension. It is based on open-source data and modern post-training techniques.

Architecture

The model is built using a combination of publicly available, synthetic, and human-created datasets, primarily in English. It features comprehensive performance enhancements for tasks like MATH, GSM8K, and IFEval. The Tülu3 models have been fine-tuned from the base model allenai/Llama-3.1-Tulu-3-8B-DPO.

Training

Tülu3 models underwent extensive training using RLVR methods, with specific hyperparameters including a learning rate of 3 × 10⁻⁷ and a discount factor of 1.0. The training involved a batch size of 512 and max token lengths of up to 2,048, among other detailed configurations. The models are designed for robust performance across various benchmarks, such as HumanEval and MATH tasks.

Guide: Running Locally

Basic Steps

1. Installation: Ensure you have the transformers library installed.

   pip install transformers
   

2. Model Loading: Use the AutoModelForCausalLM class to load the model.

   from transformers import AutoModelForCausalLM
   
   tulu_model = AutoModelForCausalLM.from_pretrained("allenai/Llama-3.1-Tulu-3-8B")
   

3. Serving the Model: Use VLLM to serve the model, ensuring the chat template is correctly formatted.

   vllm serve allenai/Llama-3.1-Tulu-3-8B --max_model_len=8192
   

Cloud GPUs

For optimal performance, consider utilizing cloud GPU services such as AWS EC2, Google Cloud Platform, or Azure, which provide powerful GPU instances suited for running large models.

License

LLAMA-3.1-TULU-3-8B is distributed under the Meta's Llama 3.1 Community License Agreement, specifically for research and educational purposes. The model's usage is subject to additional terms from third-party outputs, such as the Gemma Terms of Use and Qwen License Agreement. For comprehensive details, refer to the Responsible Use Guidelines provided by AllenAI.