RedHatAI/Llama-3.1-Nemotron-70B-Instruct-HF

TEXT GENERATIONConcurrency Cost:4Model Size:70BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 10, 2025License:llama3.1Architecture:Transformer0.0K Cold

RedHatAI/Llama-3.1-Nemotron-70B-Instruct-HF is a 70 billion parameter Llama 3.1-based instruction-tuned language model, customized by NVIDIA to enhance helpfulness in LLM-generated responses. With a 32768 token context length, it achieves strong performance on alignment benchmarks, including 85.0 on Arena Hard and 57.6 on AlpacaEval 2 LC. This model is optimized for general-domain instruction following and excels at generating coherent and helpful text.

Loading preview...

Overview

RedHatAI/Llama-3.1-Nemotron-70B-Instruct-HF is a 70 billion parameter instruction-tuned large language model, built upon the Llama 3.1 architecture and customized by NVIDIA. Its primary focus is to significantly improve the helpfulness of responses to user queries. This model was trained using RLHF (specifically, REINFORCE) with the Llama-3.1-Nemotron-70B-Reward and HelpSteer2-Preference prompts datasets, starting from a Llama-3.1-70B-Instruct base.

Key Capabilities & Performance

  • Enhanced Helpfulness: Customized by NVIDIA to generate more helpful and aligned responses.
  • Strong Alignment Benchmarks: Achieves 85.0 on Arena Hard, 57.6 on AlpacaEval 2 LC, and 8.98 on GPT-4-Turbo MT-Bench, outperforming several frontier models as of October 2024.
  • Robust Instruction Following: Demonstrates improved ability to correctly interpret and respond to complex instructions, such as counting characters in words without specialized prompting.
  • Large Context Window: Supports a maximum input context of 32768 tokens.

When to Use This Model

This model is particularly well-suited for applications requiring:

  • General-domain instruction following: Where helpful, coherent, and factually-correct responses are critical.
  • Chatbot and conversational AI: Its alignment tuning makes it effective for engaging and useful interactions.
  • Applications needing high-quality text generation: Especially where response helpfulness is a key metric.

Note: While highly capable in general instruction following, this model has not been specifically tuned for specialized domains like advanced mathematics.

Finetunes
4 models