Mastering Hermes 13B for Advanced AI

Mastering Hermes 13B for Advanced AI

Key Highlights

  • Technical Advancements: Developed by Nous Research with Teknium and Emozilla, Hermes 13B features a fine-tuned model on an 8x a100 80GB DGX machine, supporting complex tasks with a 4096 sequence length.
  • Dataset and Training: Trained on 300,000+ synthetic GPT-4 outputs, curated from sources like GPTeacher and roleplay datasets, Hermes 13B excels in reliability and performance.
  • Benchmark Performance: #1 on ARC-c, ARC-e, Hellaswag, and OpenBookQA, and #2 on Winogrande among GPT-4 models, demonstrating superior benchmark performance.
  • Key Features: Offers extended response generation, minimized hallucination, and unrestricted discussion capabilities, ensuring reliability and versatility.
  • Practical Applications: Supports AI tutoring, technical documentation automation, scriptwriting, customizable chatbots, academic research, content generation, and interactive storytelling.
  • API Integration: Easily integrates with Novita AI for seamless access to Hermes 13B’s capabilities, simplifying development and deployment.
  • Optimization and Updates: Tips on optimizing dataset quality, sequence length, and model parameters to maximize performance, with an emphasis on staying updated with the latest advancements.


Curious about the capabilities of Hermes 13B, the latest marvel from Nous Research? Wondering how it stacks up against its predecessors? Join us as we delve into the intricacies of Hermes 13B, exploring its technical specifications, training data insights, practical applications and API setup. Let’s uncover the answers to these questions and more.

Understanding Hermes 13B


Hermes 13B, or Nous-Hermes-Llama2–13b is a sophisticated language model developed by Nous Research, with significant contributions from Teknium and Emozilla in fine-tuning and dataset curation. This model has been fine-tuned on an extensive dataset of over 300,000 instructions, aiming to enhance capabilities beyond its predecessor, Hermes on Llama-1.

Technical Specifications

The fine-tuning process utilized a 4096 sequence length, indicating the model’s capacity to handle complex and lengthy inputs. The training was performed on an 8x a100 80GB DGX machine, showcasing the substantial computational resources invested in its development.

Training Data

The model was trained primarily on synthetic GPT-4 outputs, ensuring high-quality knowledge retention and task completion. The dataset is a curated collection from various sources, including GPTeacher, roleplay datasets, code instruct datasets, and unpublished materials like Nous Instruct & PDACTL.

Collaboration and Acknowledgment

The model’s development was a collaborative effort involving several key contributors and organizations (Teknium, Karan4D, Nous Research, Huemin Art, and Redmond AI), with special thanks to Redmond AI for sponsoring the compute resources.

Prompting and Interaction

The model adheres to the Alpaca prompt format, allowing users to interact with it through structured instructions and response sections.

Performance Benchmarks

The model has been evaluated across various benchmarks, including AGI-Eval, GPT-4All Benchmark Set, and BigBench Reasoning Test. Hermes 13B is currently ranked #1 on ARC-c, ARC-e, Hellaswag, and OpenBookQA, and ranked #2 on Winogrande, compared to GPT4all’s benchmarking list.

Key Features and Capabilities

Enhanced Long-Form Response Generation

The model is fine-tuned to produce extended, detailed responses. This capability is particularly useful for tasks that require comprehensive answers, such as writing, summarization, and in-depth explanations.

Reduced Hallucination Rate

“Hallucination” in the context of language models refers to the generation of plausible but factually incorrect information. The Hermes-Llama2–13b model has been fine-tuned to minimize this, ensuring that the responses are more reliable and accurate.

Absence of Censorship Mechanisms

Unlike some models that incorporate content censorship, Hermes-Llama2–13b does not have built-in restrictions on the topics it can discuss. This allows for more open dialogue and reduces the likelihood of the model avoiding certain subjects.

High-Quality Dataset Utilization

The model was trained on a curated dataset derived from high-quality synthetic GPT-4 outputs, ensuring a robust foundation in knowledge and task execution. The diversity of the dataset contributes to the model’s versatility and effectiveness across various domains.

Practical Applications of Hermes 13B

AI Tutoring System Development

Develop a personalized e-learning platform that leverages Hermes 13B to generate dynamic, interactive lesson plans and provides detailed explanations tailored to individual learners’ needs.

Technical Documentation Automation

Create tools that automate the generation of technical documentation, such as API docs, system architecture diagrams, and user guides, using Hermes 13B’s understanding of complex technical concepts.

Scriptwriting API for Creative Industries

Build an API service for scriptwriters that uses Hermes 13B to generate dialogues, plot summaries, and character descriptions for various media formats, enhancing creative workflows.

Customizable Chatbot Framework

Design a framework for building customer service chatbots that can be easily customized using Hermes 13B’s conversational AI capabilities, allowing for domain-specific interactions.

Academic Writing and Research Assistant

Implement an AI assistant for researchers that uses Hermes 13B to draft academic papers, generate literature reviews, and suggest research directions based on existing scholarly works.

Knowledge Synthesis Engine

Develop a system that synthesizes information from various domains, creating comprehensive reports or providing cross-disciplinary insights using Hermes 13B’s ability to understand and integrate diverse topics.

Digital Marketing Content Generator

Create a content generation tool for marketers that leverages Hermes 13B to produce engaging marketing materials, social media posts, and advertising copy that align with brand guidelines.

Interactive Storytelling Platform

Develop a platform for interactive storytelling in games or other media, where Hermes 13B can craft branching narratives and character interactions based on user choices.

Setting Up Hermes 13B LLM API

Step 1: Register an Account

Navigate to the Novita AI website and click the “Log In” button found in the top menu. Currently, you can sign in using either your Google or GitHub account. Upon logging in, you will be awarded $0.5 in Credits for free!

Step 2: Generate an API Key

To authenticate with the API, include a Bearer Token in the request header (e.g., -H “Authorization: Bearer ***”). We will provide you with a new API key. 

You can also create your own key by selecting “Add new key”.

Step 3: Execute an API Call

With just a few lines of code, you can make an API call and utilize the capabilities of Hermes 13B and other advanced models:

from openai import OpenAI

client = OpenAI(
    # Get the Novita AI API Key by referring:
    api_key="<YOUR Novita AI API Key>",
model = "nousresearch/nous-hermes-llama2-13b"
completion_res = client.completions.create(
    prompt="A chat between a curious user and an artificial intelligence assistant".
    stream = True, # or False
    max_tokens = 512,

Similar Models to Hermes 13B

Mythomax-l2–13b on Novita AI

The concept behind Mythomax-l2–13b is that each layer consists of multiple tensors, each tasked with specific functions. By leveraging MythoLogic-L2’s strong comprehension abilities as the input and Huginn’s extensive writing skills as the output, the resulting model excels in both areas.

Hermes-2-pro-llama-3–8b on Novita AI

Hermes 2 Pro is an enhanced and retrained iteration of Nous Hermes 2. It features an updated and refined version of the OpenHermes 2.5 Dataset, along with a newly created Function Calling and JSON Mode dataset developed internally.

Openhermes-2.5-mistral-7b on Novita AI

OpenHermes 2.5 Mistral 7B is a state of the art Mistral Fine-tune, a continuation of OpenHermes 2 model, which trained on additional code datasets.

Maximizing Model Performance

To get the most out of Hermes 13B, it’s really important to know how to tweak it just right. Here are a few pointers that can help you shine:

Ensuring a High-Quality and Diverse Dataset

With your dataset, make sure it’s varied and top-notch for tweaking purposes. Throw in instructions from various fields so the model gets better at understanding and coming up with replies.

Optimizing Sequence Length

When looking at sequence length, try sticking to 2000 or less for best results. Don’t be afraid to test different lengths though; this could help you find a sweet spot between how well it performs and how long its responses are.

Fine-Tuning Model Parameters

For model parameters, adjust them based on what you need specifically. Tinker with settings like temperature and repetition penalty to fine-tune creativity levels and response consistency.

Staying Updated

On updates: always stay in the loop with Hermes 13B’s latest versions by regularly checking for new releases or fixes which might enhance performance even further.

By keeping these tips in mind and constantly improving your approach towards fine-tuning, squeezing every bit of potential out of Hermes 13B for sophisticated AI tasks becomes possible.


In conclusion, Hermes 13B, developed by Nous Research in collaboration with others, represents a significant leap forward in AI language models. 

We’ve explored its robust technical specifications, including its capacity for handling lengthy inputs and the extensive computational resources behind its fine-tuning. The model’s training on a curated dataset of synthetic GPT-4 outputs ensures high-quality performance across various domains, reflected in its top rankings on many benchmarks. Moreover, its applications span from AI tutoring systems to digital marketing content generation, showcasing its versatility and real-world utility.

As you consider leveraging Hermes 13B for your own projects, remember the importance of optimizing dataset quality, sequence length, and model parameters to maximize its potential. Stay updated with the latest advancements to continually enhance its performance.

Frequently Asked Questions

1. What Are the Differences Between Hermes 13B and Previous Versions?

First, Hermes 13B has got a better way of churning out longer responses. Second, You’ll notice it makes up stuff (hallucinates) less often. Third, it doesn’t have those OpenAI rules limiting what it can say or do. Moreover, across all sorts of language tasks, it just performs better. Last, they’ve trained it on an even bigger mix of data this time.

2. How to download Hermes 13B?

You can download the Hermes 13B model for free on Hugging Face.

Novita AI is the all-in-one cloud platform that empowers your AI ambitions. With seamlessly integrated APIs, serverless computing, and GPU acceleration, we provide the cost-effective tools you need to rapidly build and scale your AI-driven business. Eliminate infrastructure headaches and get started for free — Novita AI makes your AI dreams a reality.
Recommended Reading
Introducing Openhermes 2.5: Comprehending the Power of Messenger of Gods
Exploring MythoMax-L2–13B: Advantages & Limits