Hugging Face MachiNoDolphin-Qwen2.5-72b Overview
MachiNoDolphin-Qwen2.5-72b is a 72.7 billion parameter large language model developed by KaraKaraWitch. It is a sophisticated merge of multiple pre-trained models, built upon the robust Qwen/Qwen2.5-72B base using the TIES merge method. This merging technique combines the strengths of several models, including sophosympatheia/Evathene-v1.0, KaraKaraWitch/SteyrCannon-0.2-Qwen2.5-72b, m8than/banana-2-b-72b, shuttleai/shuttle-3, and ZeusLabs/Chronos-Platinum-72B.
Key Capabilities & Features
- Merged Architecture: Leverages the TIES method to integrate diverse model capabilities into a single, powerful 72.7B parameter model.
- Multilingual Support: Designed to handle a broad spectrum of languages, including Chinese (zho), English (eng), French (fra), Spanish (spa), Portuguese (por), German (deu), Italian (ita), Russian (rus), Japanese (jpn), Korean (kor), Vietnamese (vie), Thai (tha), and Arabic (ara).
- ChatML Prompting: Utilizes the ChatML format for conversational interactions, aligning with common practices for Qwen 2.5 chat models.
- "Mostly Uncensored" Output: Aims to provide largely unrestricted responses, with suggestions for adjusting the system prompt for the rare instances of content filtering.
Recommended Usage
This model is suitable for general-purpose chat applications where a large, versatile, and less restrictive language model is desired. Its merged nature suggests a broad range of potential applications, from creative writing to complex reasoning tasks, benefiting from the combined knowledge of its constituent models. Users are encouraged to experiment with the suggested inference settings (Temp: 1.3-1.2, MinP: 0.05, TopA: 0.2, RepPen: 1.05) for optimal performance.