NVIDIA Releases Nemotron-4 340B, Ranks First on Hugging Face for Synthetic Data Generation and LLM Training

Jun 14, 2024, 05:00 PM

NVIDIA has announced the release of Nemotron-4 340B, a family of open-source models designed for synthetic data generation and training large language models (LLMs) for commercial applications. The models include Base, Instruct, and Reward variants and are optimized for use on 8 NVIDIA H100 GPUs. Nemotron-4 340B is competitive across many tasks and ranks first on the Hugging Face RewardBench leaderboard. The models are intended for use in various industries, including healthcare, finance, manufacturing, and retail. NVIDIA has implemented a permissive license, allowing developers to own any derivative and model output. The model, tested under the codename 'june-chatbot,' surpasses Llama-3-70b on hard benchmarks like Arena-Hard-Auto and is optimized for NVIDIA NeMo and TensorRT-LLM.