NousResearch unveiled their latest AI model, Hermes-2-Theta-Llama-3-70B, a unique blend of Hermes 2 Pro by NousResearch and Llama-3 Instruct from Meta. Developed by Charles Goddard and Arcee AI using MergeKit technology and further honed using Reinforcement Learning from Human Feedback (RLHF), the novelty of this AI model lies in its ability to unify the strengths of its progenitors, delivering coherent, contextually precise text.
The salient feature of Hermes-2-Theta-Llama-3-70B is its expertise with structured outputs and function calling. It uses ChatML for prompt formatting which aids in the construction of steerable multi-turn dialogues, a key aspect for interactive chatbots or virtual assistants. Moreover, the ability to produce structured outputs is ameliorated by training on specific system prompts, which cultivates JSON-formatted responses from the model. The model can parse function calling formats to generate API calls, interpret the responses and revert with arranged data. This comes handy in tasks like extracting relevant features from documents or making real-time data queries.
Benchmarking against other AI models reveals the prowess of Hermes-2-Theta-Llama-3-70B. Successfully scoring high in various benchmarks, such as GPT4All, AGIEval, and BigBench, the model outshines others in logical reasoning, knowledge-based queries and in generating factually accurate responses, as seen from its performance in the TruthfulQA benchmark.
With a broad scope of applications, the model can alternate between a variety of characters such as an anime catgirl skilled in coding and hacking to a 17th-century alchemist on the hunt for the philosopher’s stone. These functions are pivotal in creative writing, interactive storytelling, and creating engaging virtual characters. Taking into account its proficiency in function calling and creating structured outputs, it can effectively present market data for financial analysts and pair well with existing systems via API calls for business applications.
Available on platforms such as Hugging Face and on NousResearch’s GitHub Repository, users can access this model through Inference Endpoints for dedicated usage. Quantized versions of the model are also available for apps that require fewer computational resources.
In essence, NousResearch’s Hermes-2-Theta-Llama-3-70B acts as a revolutionary model for text generation, structured outputs, and function calling. Its myriad applications extend from interactive storytelling and creative writing to business intelligence and finance, thereby setting a new standard in the realm of AI and text generation.
