OpenHermes-13B

OpenHermes-13B is a fine-tuned language model built on a fully open-source dataset of 242,000 entries, primarily generated by GPT-4 and curated from various open AI datasets. It was developed to improve text generation by removing AI disclaimers and refusals, aiming for clearer and more direct communication outputs. The model incorporates data from multiple contributors, including Teknium, WizardLM Team, Microsoft, and others, ensuring a diverse and comprehensive training base.

This model is designed for developers and researchers who want an open-source alternative for advanced text generation tasks without the typical AI safety disclaimers. It supports integration with popular machine learning libraries like Transformers and can be deployed locally or via Docker, making it accessible for experimentation and production use.

OpenHermes-13B offers transparency in its training process, with public access to its WANDB project logs and detailed benchmark results. It shows competitive performance on benchmarks like GPT4All and BigBench, with slight improvements over similar models, although it has some trade-offs in AGI-Eval scores.

Technically, it was trained using multi-GPU setups with Adam optimizer and cosine learning rate scheduling, emphasizing reproducibility and open collaboration. The model is compatible with various inference tools and quantizations, allowing use in environments like llama.cpp and LM Studio.

Overall, OpenHermes-13B provides a valuable resource for those seeking a powerful, open-source language model fine-tuned for instruction following and text generation without restrictive AI disclaimers, suitable for research, development, and deployment in diverse AI applications.