Introduction to NousResearch’s Hermes 3

A New Standard in AI Models

Hermes 3 by NousResearch Source

Author

• Tohfa Siddika Barbhuiya (ORCID: 0009–0007–2976–4601)

Introduction

Instruct (or “chat”) tuned models have become the primary way in which most people interact with large language models. As opposed to “base” or “foundation” models, instruct-tuned models are optimised to respond to imperative statements.

NousResearch’s Hermes 3 is a breakthrough in Generalist Instruction Models, offering advancements in Language Models and versatile AI Capabilities. Hermes 3 contains advanced long-term context retention and multi-turn conversation capability, complex roleplaying and internal monologue abilities as claimed by the developers. Hermes 3 training data aggressively encourages the model to follow the system and instruction prompts exactly and in an adaptive manner. Hermes 3 was created by fine-tuning Llama 3.1 8B, 70B and 405B, and training on a dataset of primarily synthetically generated responses. These models are not only powerful but also uniquely aligned to follow system and instruction prompts with unparalleled precision and neutrality. The model boasts comparable and superior performance to Llama 3.1 while unlocking deeper capabilities in reasoning and creativity. This article deep dives into the innovative features of Hermes 3, its advanced capabilities, and its impact on the field of AI. This Generalist AI System showcases Intelligent Conversational Agents and Adaptive AI Algorithms, enabling Personalised AI Assistance and AI-driven Task Completion. The advent of large language models like ChatGPT has revolutionised how we interact with AI, introducing the “chat” paradigm where AI assumes the role of a conversational assistant. This shift towards chat-tuned models has made AI more accessible and user-friendly, allowing it to respond naturally to human inquiries and commands. However, the true potential of large language models extends beyond simple conversation. Hermes 3 is not just another chatbot. This model is a versatile assistant capable of complex tasks.

Some key concepts

• Generalist Instruction Models: Generalist Instruction Models are AI systems designed to understand and execute a wide variety of tasks based on given instructions. Unlike specialised models that are fine-tuned for specific tasks, generalist models can adapt to multiple contexts, allowing them to perform diverse functions like answering questions, generating creative content, engaging in dialogue, or executing complex commands across different domains. These models are trained on broad datasets and are optimised to interpret.
• Multi-Turn Conversation Capability: The AI can remember what was discussed earlier and continue the conversation naturally. Example: You ask the AI, “What’s the weather like today?” After it tells you, you say, “Should I take an umbrella?” The AI remembers the weather details and advises you accordingly.
• Internal Monologue Abilities: The AI thinks through a problem internally before giving you an answer. Example: You ask the AI, “What’s the best gift for a 10-year-old?” The AI considers factors like age, interests, and current trends before suggesting a popular toy.

Model Overview and Key Features

Hermes 3, particularly its flagship 405B model, is designed to excel in a variety of complex tasks. The model’s advanced reasoning and creative capabilities are complemented by several key features that set it apart from its predecessors:

• Advanced Agentic Capabilities: Agentic capabilities imply that the system is not just passively following instructions, but rather actively engaging in decision-making processes, adapting to new situations, and possibly even initiating actions to achieve specific goals. Hermes 3 demonstrates refined autonomy, making it capable of handling complex, goal-oriented tasks with minimal supervision.
• Enhanced Roleplaying and Reasoning: The model excels in maintaining coherent, contextually appropriate interactions over multiple turns, making it ideal for applications in virtual assistants and interactive storytelling.
• Improved Long-Context Coherence: Hermes 3 maintains the flow of conversation and context over extended dialogues, which is crucial for long-form content creation and analysis.
• Structured Output and Function Calling: With advanced capabilities in generating structured outputs like JSON and invoking external tools, Hermes 3 is a versatile tool for developers looking to integrate AI into existing systems.

Innovative Features of Hermes 3

Hermes 3 introduces several cutting-edge features that enhance its ability to solve complex problems:

• Structured Output with XML Tags: The model can generate well-organised outputs using XML tags, facilitating structured communication.
• Intermediate Processing using Scratchpads: Hermes 3 uses scratchpads for intermediate steps, ensuring transparency and accuracy in decision-making. That is, the AI system records and displays its intermediate steps during decision-making or problem-solving processes. These “scratchpads” act like temporary workspaces where the AI can jot down thoughts, calculations, or partial results before reaching a final conclusion.
• Transparent Decision-Making through Internal Monologues: The model provides insight into its reasoning process through internal monologues, making its decision-making more understandable.
• Step-Labeled Reasoning and Planning: The model labels each step in its reasoning and planning process, offering a clear roadmap of its problem-solving approach.

One of the key features of Hermes 3 is its ability to make informed judgments and assess the quality of generated text through improved judgment and reward modelling. This allows Hermes 3 to evaluate and refine its outputs with a nuanced understanding, potentially enabling more effective fine-tuning and iterative improvements in language models. An example of this capability can be seen in multi-turn evaluations, where the model demonstrates refined decision-making across extended interactions.

In addition, Hermes 3 incorporates several advanced agentic capabilities designed to enhance multi-step problem-solving and transparency in reasoning. The model utilises XML tags for structured output and implements scratchpads to facilitate intermediate processing. These tools enable Hermes 3 to generate internal monologues that provide insights into its decision-making process and utilise step-labeled reasoning for systematic planning. The model was trained on various reasoning tasks using specialised tokens such as <SCRATCHPAD>, <REASONING>, <INNER_MONOLOGUE>, <PLAN>, and others, significantly improving its ability to tackle complex tasks and clearly communicate its methodology.

For instance, in the domain of software development, Hermes 3 excels in generating complex, functional code snippets across various programming languages. It can also provide detailed explanations and documentation, showcasing a deep understanding of coding paradigms and design patterns. This makes it an invaluable tool for developers, particularly in tasks like planning and implementing a Discord chatbot, where agentic coding generation is critical.

Data Mixture and Training Recipe

The Hermes 3 dataset is a carefully curated collection of 390 million tokens, covering a wide range of domains. This diverse mixture includes:

• 270M Response Tokens (69%): Ensuring robust responses across different scenarios.
• 120M Instruction Tokens (31%): Fine-tuning the model’s ability to follow complex instructions.

The training process of Hermes 3 is executed in two phases:

• Supervised Fine-Tuning (SFT): This phase optimises the model’s response quality and accuracy.
• Direct Preference Optimization (DPO): DPO offers further refinement, particularly enhancing the performance of the 8B model.

Hermes 3 performance benchmarks

Hermes 3 has undergone rigorous testing against a variety of benchmarks, showcasing its competitive, if not superior, capabilities compared to Llama-3.1 Instruct models. The Hermes 3 Technical Report provides detailed comparisons and analyses, highlighting the model’s strengths and areas for improvement.

Hermes 3 performance evaluation Source

Prompt Format and Inference

Hermes 3 employs the ChatML prompt format, which facilitates structured and intuitive multi-turn chat dialogues. Key aspects include:

• System Prompts: These allow for high steerability, enabling users to define roles, stylistic choices, and interaction rules.
• Function Calling and Structured Outputs: The model’s training on specific system prompts supports function calling, making it an effective tool for complex data retrieval and processing.

With its massive 405 billion parameters, Hermes 3 requires significant computational resources. NeuralMagic’s FP8 quantisation method reduces the model’s VRAM requirements, making it more accessible for large-scale inference tasks.

Hermes 3 — Llama-3.1 8B — Install and Test Locally

I have used Anaconda and Jupyter notebook for this project. Open your Anaconda prompt.

1. Create a new Conda environment named “hermes” with Python 3.11

conda create -n hermes python=3.11 -y

2. Activate the “hermes” Conda environment

conda activate hermes

3. Install PyTorch in the “hermes” environment

conda install pytorch torchvision torchaudio cudatoolkit=11.8 -c pytorch

4. Upgrade the Transformers library to the latest version

pip install --upgrade transformers

5. Install the Accelerate library and Hugging Face Hub

pip install accelerate huggingface hub

6. Install Jupyter Notebook in the “hermes” environment

conda install jupyter -y

7. Uninstall the charset_normalizer package if it is installed.

pip uninstall charset_normalizer -y

8. Reinstall the charset_normalizer package to ensure you have the correct version.

pip install charset_normalizer

9. Launch Jupyter Notebook to start creating and managing your notebooks.

jupyter notebook

10. Now create a new notebook and start working

12. Import essential libraries for working with transformers and PyTorch

import transformers 
import torch 
from IPython.display import Markdown,display

13. Set up a text generation pipeline using the Hermes-3 model from Hugging Face Transformers

import transformers

model_id = "NousResearch/Hermes-3-Llama-3.1-8B"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto"
)

14. Generate a response from the Hermes-3 model based on a user quiz query and display the output

messages = [
    {"role": "system", "content": "You are helpful assistant!"},
    {"role": "user", "content": "which is the largest sand island in the world and in which country is it?" }
]

outputs = pipeline(messages, max_new_tokens=1024)
output_text = outputs[0]["generated_text"][-1]['content']
display(Markdown(output_text))

Response:

Image generated from the desktop of Tohfa Siddika Barbhuiya

Solve a math problem using the Hermes-3 model and display the solution

prompt = """
A notebook and pen costs $2.20. The notebook costs $1.00 more than the pen. How much does the pen costs"
"""
messages = [
{"role": "system", "content": "You are helpful assistant!"},
{"role": "user", "content": prompt }
]

outputs = pipeline(messages, max_new_tokens=1024)
output_text = outputs[0]["generated_text"][-1]['content']
display(Markdown(output_text))

Response:

Image generated from the desktop of Tohfa Siddika Barbhuiya

Training Focus Areas

The training of Hermes 3 focused on several key areas to enhance its capabilities:

• Agentic Capabilities: Improved autonomy in complex scenarios.
• Roleplaying: Enhanced understanding and maintenance of different personas.
• Reasoning: Significant advancements in logical problem-solving.
• Multi-turn Conversation: Improved coherence and context across extended dialogues.
• Long Context Coherence: Better relevance maintenance over longer text passages.

Conclusion

Hermes 3 is a groundbreaking advancement in the field of generalist instruction models. With its cutting-edge features, enhanced tool use capabilities, and robust training, Hermes 3 sets a new standard for AI models in various domains. Whether you’re a researcher, developer, or AI enthusiast, the Hermes 3 AI model offers powerful tools and functionalities that can drive innovation and elevate your projects to new heights. For more details on its capabilities, benchmarks, and usage, refer to the Hermes 3 Technical Report.

References

• Nous Research, “Hermes 3,” Nous Research. [Online]. Available: https://nousresearch.com/hermes3/. [Accessed: 21-Aug-2024].
• Nous Research, “Hermes 3 Collection,” Hugging Face. [Online]. Available: https://huggingface.co/collections/NousResearch/hermes-3-66bd6c01399b14b08fe335ea?source=post_page-----cdbe2cb27a2d--------------------------------. [Accessed: 21-Aug-2024].
• Nous Research, “Hermes 3 Technical Report,” Aug. 2024. [Online]. Available: https://nousresearch.com/wp-content/uploads/2024/08/Hermes-3-Technical-Report.pdf. [Accessed: 21-Aug-2024].