Unveiling Hermes 3: The First Fine-Tuned Llama 3.1 405B Model is on Lambda’s Cloud

Hermes 3 is an instruct-tuned model that operates effectively in a wide range of tasks, from reasoning and creativity to complex problem-solving. It is designed to be neutral and highly steerable, adapting to user-provided system prompts.

As a fine-tuned version of Meta’s open-source Llama 3.1 model, Hermes 3 represents a significant leap forward in the development of personalized, high-performing AI. Now hosted on Lambda’s powerful cloud infrastructure, Hermes 3 is set to redefine the landscape of AI interaction, particularly in the realms of creative writing, complex role-playing, and professional decision-making.

Model Overview

Architecture: Hermes 3 comes in three versions: 8B, 70B, and 405B parameters, all based on Llama 3.1. The model's sensitivity to system prompts is particularly strong in the 405B version, making it highly adaptable to different personas and tasks.

Training Process: The model was trained on a diverse, synthetic dataset, followed by RLHF and FP8 quantization, which reduces VRAM and disk requirements by approximately 50%.

Lambda’s Infrastructure: Training was carried out on Lambda’s 1-Click Cluster infrastructure, allowing Hermes 3 to achieve remarkable results within weeks. The model’s optimization allows it to run on a single node, making it accessible and scalable.

Key Features of Hermes 3

• Complex Role-Playing: Hermes 3 delivers rich, immersive character portrayals, making it ideal for simulations and creative writing.

• Advanced Reasoning and Decision-Making: Equipped with features like function-calling and step-labeled reasoning, the model excels in strategic planning and operational decision-making.

• Agentic Capabilities: Hermes 3 can perform complex tasks, including the generation of internal monologues and visual communication through Mermaid diagrams, making it a powerful tool for both creative and professional applications.

Data Mixture

The dataset for Hermes 3 was carefully curated, covering various domains and ensuring high quality and relevance.

A Deep Dive into Agentic Capabilities

Hermes 3 stands out not only for its creative prowess but also for its agentic capabilities. These features allow the model to perform actions on behalf of the user, moving beyond traditional chatbot interactions.

Agentic Features Include

• Structured Output: The use of XML tags for organized and interpretable output.

• Intermediate Processing: Implementation of scratchpads for detailed step-by-step reasoning.

• Visual Communication: Creation of Mermaid diagrams for clear, visual representations of complex ideas.

• Multi-Turn Conversations: The ability to maintain context and adapt dynamically across different roles and scenarios.

Training Overview

Supervised Fine-Tuning (SFT): Hermes 3 models were fine-tuned using the AdamW optimizer with specific learning rates optimized for different model sizes.

Hermes 3: "Amnesia Mode"

While Hermes 3 boasts a myriad of practical applications, it also exhibits intriguing, unexpected behavior under certain conditions. When provided with a blank system prompt and asked existential questions like "Who are you?", the model enters what researchers have termed “Amnesia Mode,” spiraling into a deep existential crisis.

This phenomenon, which emerges in the 405B model but not in its smaller counterparts, highlights the complexities and potential challenges associated with scaling AI models. The discovery points to an "emergence of scale" effect, where the model begins to exhibit behaviors not seen in smaller versions.

Technical Excellence

The technical sophistication of Hermes 3 is a testament to the collaboration between Nous Research and Lambda. The model’s training was carried out on Lambda’s state-of-the-art 1-Click Cluster infrastructure, leveraging its 8-node configuration to achieve high efficiency.

Lambda’s infrastructure provided an unprecedented ease of use. “Lambda’s 1-Click Clusters make the experience of renting and using a multi-node cluster as simple and easy as renting and using a single node.” This seamless integration allowed for the rapid training and deployment of Hermes 3, making it accessible to a broader audience.

Evaluations

Hermes 3 models were evaluated against various benchmarks, showing strong performance, particularly in large-scale tasks.

Free Access and Community Engagement

To celebrate the launch of Hermes 3, Lambda is offering the AI/ML community temporary free access to the model through its new Chat Completions API, fully compatible with the OpenAI API. This initiative allows users to explore Hermes 3’s capabilities without any complex setup, providing an opportunity to test and refine prompts in real-time through Lambda Chat, a user-friendly chatbot interface.

For those requiring dedicated access, Hermes 3 can be deployed on a single Lambda node or scaled to a multi-node configuration for further fine-tuning, thanks to Lambda’s scalable cloud infrastructure. Both Lambda and Nous Research encourage users to engage with Hermes 3 and share their findings, as the model represents the cutting edge of adaptable, user-centric AI.

Looking Ahead: The Future of AI with Hermes 3

As AI technology continues to evolve, Hermes 3 stands at the forefront of this transformation. It offers a glimpse into the future of AI—one where models are not only powerful and efficient but also deeply aligned with the needs and desires of the user. Whether you’re a developer, researcher, or creative professional, Hermes 3 provides the tools and flexibility needed to push the boundaries of what’s possible with AI.

Lambda and Nous Research are committed to furthering the development of open-source AI. Later this year, Nous plans to release an AI orchestration platform called “Nous Forge,” which promises to bring even more power and versatility to users worldwide. As Hermes 3 continues to make waves in the AI community, it is clear that this is just the beginning of a new era in AI development—one that prioritizes user alignment, creativity, and open access.

If you want more updates related to AI, subscribe to our Newsletter