Nous-Hermes-Llama2-13b: Instruction-Tuned Llama 2 Model

Nous-Hermes-Llama2-13b is a 13 billion parameter language model from Nous Research, built upon the Llama 2 architecture and fine-tuned with a 4096 token context length. The model was trained on over 300,000 instructions, primarily derived from high-quality synthetic GPT-4 outputs, ensuring robust knowledge and task completion capabilities. Key contributors to its development include Teknium and Emozilla for fine-tuning and dataset curation, with compute sponsored by Redmond AI.

Key Capabilities & Differentiators

Enhanced Instruction Following: Fine-tuned on a diverse range of GPT-4 generated datasets, including GPTeacher, Wizard LM, and Nous Instruct, for superior instruction adherence.
Reduced Hallucination & Longer Responses: Engineered to produce more coherent and extended outputs with a lower propensity for factual errors compared to its predecessors.
Unrestricted Output: Lacks OpenAI's inherent censorship mechanisms, offering greater flexibility in content generation.
Strong Benchmark Performance: Achieves high scores on various benchmarks, including a 70.0 average on the GPT4All benchmark set and 0.372 on AGIEval. It holds top positions on ARC-c, ARC-e, Hellaswag, and OpenBookQA, and second place on Winogrande among GPT4all's listed models.

Use Cases

This model is well-suited for applications requiring detailed instruction following, creative text generation, and complex reasoning. Its design makes it a strong candidate for chatbots, content creation, and tasks where nuanced understanding and extended, uncensored responses are beneficial.