Nemotron-4-340B-Reward, developed by NVIDIA, is a multi-dimensional reward model used in synthetic data generation pipelines to assist researchers and developers in building their own LLMs. Composed of the Nemotron-4-340B-Base model and a linear layer, it converts response end-of-sequence markers into five scalar values corresponding to HelpSteer2 attributes. It supports a maximum context length of 4096 tokens and can score five attributes for each assistant turn.