nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual
The nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual is a 49 billion parameter generative reward model developed by NVIDIA, built upon the Llama-3.3-Nemotron-Super-49B-v1 foundation. Fine-tuned with Reinforcement Learning, it predicts the quality of LLM-generated responses and can rank two responses given a multilingual conversation history. This model excels at judging response quality by generating reasoning traces and outputting an integer score, with higher scores indicating better quality.
Loading preview...
Model Overview
The nvidia/Llama-3_3-Nemotron-Super-49B-GenRM-Multilingual is a 49 billion parameter generative reward model (GenRM) from NVIDIA, based on the Llama-3.3-Nemotron-Super-49B-v1 architecture. It is specifically fine-tuned using Reinforcement Learning to evaluate the quality of responses generated by large language models.
Key Capabilities
- Response Quality Prediction: The model can assess the quality of a single LLM-generated response, providing an integer score where higher values indicate better quality.
- Multilingual Response Ranking: It can rank the quality of two different responses given a multilingual conversation history, outputting individual scores and a comparative ranking score.
- Reasoning Trace Generation: Before providing a score, the model generates detailed reasoning traces to explain its judgment.
- Top Performance: As of May 2025, this model, trained with HelpSteer3-Preference, is a top-performing Bradley-Terry reward model on the RM-Bench and JudgeBench leaderboards, outperforming corresponding Bradley-Terry reward models.
Use Cases
This model is ideal for developers and researchers looking to:
- Automate LLM Response Evaluation: Programmatically assess the quality of LLM outputs in various domains like chat, math, code, and safety.
- Compare LLM Performance: Rank and compare the effectiveness of different LLM responses to a given prompt.
- Enhance LLM Development Workflows: Integrate into feedback loops for fine-tuning or improving generative AI models by providing objective quality metrics.
Technical Details
The model supports a maximum context length of 128k tokens and is optimized for NVIDIA GPU-accelerated systems, including Ampere and Hopper architectures. It is recommended for deployment with vLLM 0.8.3 on systems with 2 or more 80GB GPUs.