Meta Llama 3 70B Instruct

Meta Llama 3 70B Instruct Terms of Use

This model is licensed under the Meta Llama 3 Community License Agreement.

Model Description

Meta-Llama-3-70B-Instruct is an instruct-tuned decoder-only, text-to-text model. It was trained on 15 trillion tokens of data from publicly available sources. The instruction-tuning uses supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.

Meta-Llama-3-70B-Instruct builds on the existing work done for the Llama, and Llama 2, family of models. It makes a few key improvements:

• 128K token vocabulary, allowing for language to be encoded more efficiently
• Grouped-Query Attention (GQA) for all sizes of the model
• 8,192 context window for training with cross-document masking

In their development of these models, Meta took great care to optimize for helpfulness and safety.

More details on the model can be found here.

This model is optimized through NVIDIA NeMo Framework, and is provided through a .nemo checkpoint.

Getting Started

Step 1: Fetching the model file

1. You can download the model from NGC using the CLI tool.

ngc registry model download-version nvidia/nemo/llama-3-70b-instruct-nemo:1.0

You can find the instructions to install and configure the NGC CLI tool here

Step 2: Fetch the NeMo Framework container

You can use the NeMo Framework container available on NGC which comes preloaded with all the required dependencies.

export NEMO_IMAGE="nvcr.io/nvidia/nemo:24.05"
docker pull $NEMO_IMAGE

Step 3: Run the NeMo container

Run the NeMo framework container, mounting the model.

# Full path to the extracted model directory
export MODEL_DIR=$(pwd)/llama-3-70b-instruct-nemo
docker run --gpus all -it --rm  -v ${MODEL_DIR}:/model $NEMO_IMAGE

Step 4: Follow The Desired Playbooks (Coming Soon!)

We're hard at work preparing resources to enable you to most easily use the Meta-Llama-3-70B-Instruct .nemo checkpoint!