| Property | Value |
|---|---|
| Parameter Count | 70 Billion |
| Context Length | 8,192 tokens |
| License | Llama 3 Community License |
| Base Model | Llama 3 70B Base |
| Paper | HelpSteer2 Paper |
What is Llama3-70B-SteerLM-RM?
Llama3-70B-SteerLM-RM is a sophisticated reward model built on the Llama 3 70B architecture, designed to evaluate AI responses across multiple dimensions. Unlike traditional reward models that provide a single score, this model assesses responses on five distinct attributes: helpfulness, correctness, coherence, complexity, and verbosity, each rated on a scale of 0 to 4.
Implementation Details
The model is implemented using NVIDIA's NeMo-Aligner framework and trained on the HelpSteer2 dataset. It achieves impressive performance on the RewardBench leaderboard, scoring 88.8% overall and particularly excelling in safety evaluations with a 92.8% score.
- Built with NVIDIA NeMo Framework for scalable training
- Supports both multi-aspect and single-scalar reward outputs
- Implements efficient data and model parallelism
- Compatible with the entire NeMo ecosystem
Core Capabilities
- Multi-dimensional response evaluation across 5 key attributes
- High-performance safety assessment capabilities
- Flexible deployment options through NeMo-Aligner
- Support for both float and integer-based attribute scoring
- 8,192 token context window for comprehensive conversation analysis
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its multi-aspect evaluation approach, providing granular insights into response quality across five different dimensions, rather than just a single score. It's also one of the top-performing open-source reward models available.
Q: What are the recommended use cases?
The model is ideal for response quality evaluation, dialogue system training, and SteerLM training. It can be used both as a multi-aspect reward model for detailed analysis or as a conventional single-score reward model using provided weight configurations.