Nous Research Unveils Hermes 3 AI Models
Nous Research has released a new family of AI models called Hermes 3, which are notable for their high level of control and neutrality, distinguishing them from other models that may have ethical restrictions. Hermes 3 models are available in three sizes—8, 70, and 405 billion parameters—and are built on Meta's open-source Llama 3.1. These models are designed to closely follow user commands and can adapt to different worldviews as specified by users.
The Hermes 3 models excel in tasks such as reasoning, reward modeling, and generating structured outputs, including XML tags. Additionally, they are capable of generating internal monologues for transparent decision-making and creating visual content like Mermaid diagrams. The models were trained through a two-step process, involving supervised fine-tuning and direct preference optimization, utilizing nearly 400 million tokens in the initial phase.
In testing, Hermes 3 models performed well, leading among open-source models in benchmarks such as ARC, BoolQ, HellaSwag, IFEval, and Winogrande. The models were trained on a mix of reasoning tasks and creative applications, such as role-playing and writing. They also have the ability to use external tools and retrieve information from documents through Retrieval Augmented Generation (RAG), enhancing their ability to provide accurate and relevant answers.
Available on Hugging Face, the Hermes 3 models are gaining recognition for their versatility across a range of applications, from business decision-making to creative tasks. Experts view Hermes 3 as a significant player in the AI landscape, reflecting a broader industry trend toward open-source AI models that offer flexibility and cost-efficiency for businesses.
Key Takeaways
- Nous Research releases Hermes 3, a family of AI language models based on Meta's Llama 3.1.
- Hermes 3 models come in 8, 70, and 405 billion parameters, designed for high controllability and neutral alignment.
- The models excel in tasks like reasoning and structured output, achieving top scores in public benchmarks.
- Training involved supervised fine-tuning and direct preference optimization, using nearly 400 million tokens.
- Hermes 3 models are available for use on Hugging Face, supporting external tools and document-based information retrieval.
Analysis
Nous Research's Hermes 3 AI models, based on Meta's Llama 3.1, could disrupt industries requiring precise AI interactions. Their high controllability and neutral alignment enhance applications in finance, healthcare, and legal sectors. The availability on Hugging Face broadens their accessibility, potentially boosting Nous Research's market share. Competitors may accelerate innovations to match Hermes 3's capabilities. Long-term, these models could standardize AI behavior, influencing ethical AI development globally.
Did You Know?
- Hermes 3 Models:
- Explanation: Hermes 3 is a series of advanced AI language models developed by Nous Research, based on Meta's open-source Llama 3.1 architecture. These models are notable for their high level of controllability and neutrality, meaning they can adhere strictly to user commands without ethical constraints that might limit other AI models. They are available in three sizes—8, 70, and 405 billion parameters—each designed to handle complex tasks such as reasoning, reward modeling, and structured output using XML tags.
- Direct Preference Optimization (DPO):
- Explanation: Direct Preference Optimization is a training technique used in the development of the Hermes 3 models. Unlike traditional methods that rely on supervised learning followed by reinforcement learning from human feedback, DPO directly optimizes the model's responses based on human preferences expressed through a reward model. This approach streamlines the training process and enhances the model's ability to align with user intentions more effectively.
- Retrieval Augmented Generation (RAG):
- Explanation: Retrieval Augmented Generation is a technique that enhances the capabilities of AI models like Hermes 3 by allowing them to access and utilize external information sources, such as documents or databases, during the generation of responses. This means the models can provide more accurate and contextually relevant answers by pulling in relevant information from these external sources, thereby augmenting their natural language generation abilities.