NVIDIA Nemotron

Nemotron-4-340B-Instruct, a fine-tuned version of Nemotron-4-340B-Base optimized for synthetic data generation and chat use-cases, was pre-trained on 9 trillion tokens from diverse texts, languages, and coding languages, and further aligned using 20K human-annotated samples and 98% synthetic data through SFT, DPO, and RPO, with a detailed synthetic data generation pipeline outlined in the technical report /start