Hugging Face Overview
Dans-PersonalityEngine-V1.3.0-12b Overview
Dans-PersonalityEngine-V1.3.0-12b is a 12 billion parameter language model developed by PocketDoc, built upon the mistralai/Mistral-Nemo-Base-2407 architecture. It has been extensively fine-tuned on over 50 specialized datasets, enabling it to handle a diverse range of tasks.
Key Capabilities & Features
- Versatile Performance: Designed for both creative applications (e.g., roleplay, co-writing) and technical tasks (e.g., code generation, tool use, complex reasoning).
- Multilingual Support: Version 1.3.0 introduces support for 10 languages, including Arabic, Chinese, English, French, German, Hindi, Japanese, Korean, Portuguese, and Spanish. English remains the primary language for optimal performance.
- Extended Context Window: Features a context length of 32768 tokens, with degraded recall up to 131072 tokens.
- Optimized Prompting: Utilizes a custom "DanChat-2" format with special tokens for each role, aiming to reduce biases and enhance adaptability compared to standard ChatML.
- Training Details: Trained using Axolotl on 8x H100 GPUs for 27 hours, with resources provided by Prime Intellect and Kalomaze.
Recommended Use Cases
- Creative Content Generation: Ideal for applications requiring imaginative text, role-playing scenarios, and collaborative writing.
- Technical Problem Solving: Suitable for code generation, leveraging external tools, and tackling intricate reasoning challenges.
- Multilingual Applications: Can be employed in projects requiring understanding and generation across its 10 supported languages, with best results in English.