ANITA-NEXT-24B-Dolphin-Mistral-UNCENSORED-ITA Overview

This model, developed by Ph.D. Marco Polignano and the SWAP Research Group, is a 24 billion parameter "Thinking Model" within the ANITA (Advanced Natural-based interaction for the ITAlian language) family. It is a fine-tuned version of dphn/Dolphin-Mistral-24B-Venice-Edition, which itself is based on the Mistral architecture. A key characteristic is its uncensored nature, making it a multilingual model supporting both English and Italian, with the explicit warning that it may exhibit dangerous, unethical, or offensive behaviors.

Key Capabilities

• Uncensored Responses: Designed to generate content without typical safety filters, suitable for specific research applications.
• Multilingual Support: Proficient in both English and Italian.
• Mistral Architecture: Leverages the efficient and capable Mistral base.
• Extended Context Window: Supports a context length of 32768 tokens, though performance may degrade after 40k tokens.
• QLoRA and DPO Fine-tuning: Utilizes QLoRA 4-bit for supervised fine-tuning on instruction-based datasets and DPO over mlabonne/orpo-dpo-mix-40k for alignment.

Good For

• Research Purposes: Specifically released for research into uncensored model behaviors and capabilities.
• Italian Language Use Cases: Part of the ANITA project aimed at improving NLP for the Italian language.
• Exploring Model Limitations: Useful for studying the outputs of models without strict safety alignments.