dolphin-2.2.1-mistral-7b is an instruction-tuned, uncensored language model developed by Eric Hartford and Cognitive Computations. It is based on the Mistral-7B-v0.1 architecture and fine-tuned on a diverse combination of datasets, including an open-source implementation of Microsoft’s Orca dataset, Airoboros, WizardLM, and a curated subset of Samantha data.
The model is specifically designed for high compliance and empathy, allowing it to provide personal advice and maintain an emotional tone during interactions. As an uncensored model, it has been filtered to remove alignment and bias instructions, meaning it will respond to a wide range of queries, including those that standard aligned models might refuse.
Dolphin 2.2.1 utilizes the ChatML prompt format to facilitate structured multi-turn conversations. It was released as a checkpoint update to address overfitting issues observed in previous versions, specifically to reduce unrequested Chain-of-Thought (CoT) responses and improve general instruction following. The model is released under the Apache-2.0 license, making it suitable for both commercial and non-commercial applications.