| Property | Value |
|---|---|
| Parameter Count | 70.6B |
| License | Llama 3.1 |
| Base Model | meta-llama/Llama-3.1-70B-Instruct |
| Paper | HelpSteer2-Preference Paper |
| Training Method | RLHF (REINFORCE) |
What is Llama-3.1-Nemotron-70B-Instruct-HF?
This is NVIDIA's enhanced version of Meta's Llama-3.1 model, specifically optimized to provide more helpful and accurate responses. As of October 2024, it achieves state-of-the-art performance on multiple benchmarks, including Arena Hard (85.0), AlpacaEval 2 LC (57.6), and MT-Bench (8.98), surpassing models like GPT-4 and Claude 3.5 Sonnet.
Implementation Details
The model builds upon Llama-3.1-70B-Instruct using RLHF techniques with NVIDIA's HelpSteer2 dataset. It's implemented in BF16 precision and requires at least 2x 80GB GPUs for inference.
- Trained using REINFORCE with HelpSteer2-Preference prompts
- Supports up to 128k input tokens
- Generates up to 4k output tokens
- Compatible with NVIDIA Ampere, Hopper, and Turing architectures
Core Capabilities
- Superior helpfulness in general-domain instruction following
- Enhanced factual accuracy and response coherence
- Customizable complexity and verbosity in responses
- Efficient handling of counting and basic reasoning tasks without specialized prompting
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its optimization for helpfulness through NVIDIA's HelpSteer2 dataset and RLHF training, achieving top performance across multiple benchmarks while maintaining natural and coherent responses.
Q: What are the recommended use cases?
The model excels in general-purpose instruction following, conversation, and tasks requiring helpful and accurate responses. However, it's not specifically tuned for specialized domains like mathematics.