Overview

MarinaraSpaghetti/NemoReRemix-12B is a 12 billion parameter language model, a merge of several pre-trained models including Mistral-Nemo-Instruct, Sao10K_MN-12B-Lyra-v1, shuttle-2.5-mini, migtissera_Tess-3-Mistral-Nemo, and anthracite-org_magnum-12b-v2. It was created using the della_linear merge method with Mistral-Nemo-Base-2407 as the base model. This iteration focuses on improving storytelling and roleplay capabilities while also serving as a general assistant.

Key Capabilities

• Enhanced Storytelling & Roleplay: Specifically improved for generating consistent prose and handling roleplay scenarios without formatting issues.
• General Assistant Functionality: Capable of performing general conversational tasks.
• Improved Intelligence: Benefits from the integration of Migtissera's Tess model, contributing to smarter responses.
• Consistent Formatting: Addresses previous issues with asterisk-based formatting, ensuring cleaner output.
• Flexible Instruction Format: Supports ChatML and theoretically Mistral Instruct formats, with a note on adding <|im_end|> to custom stopping strings.

Good For

• Creative Writing: Ideal for generating narratives and engaging in detailed storytelling.
• Interactive Roleplay: Suitable for applications requiring dynamic and consistent character interactions.
• Conversational AI: Can be used for general-purpose assistant tasks where intelligent and well-formatted responses are crucial.

Recommended Settings

For optimal performance, the creator recommends running with Temperature 1.0-1.2, 0.1 Top A or 0.01-0.1 Min P, and 0.8/1.75/2/0 DRY. Lower temperatures below 1.0 are also supported. Specific SillyTavern settings are available via a provided Hugging Face link.