NGalrion/MarinaraSpaghetti-NemoMix-Unleashed-12B-chat

Model Overview

NGalrion/MarinaraSpaghetti-NemoMix-Unleashed-12B-chat is a 12 billion parameter language model merge, developed by NGalrion. This model is a result of extensive efforts to improve ChatML merges, focusing on reducing repetition, particularly in longer contexts. It leverages a Mistral-Nemo-Base architecture and incorporates components from several other models, including those by Intervitens, Sao10K, and Nbeerbower, merged using the della_linear method.

Key Capabilities

• Reduced Repetition: Designed to be less prone to repetitive outputs, especially when handling higher context lengths.
• Extended Context: Supports a substantial context window of 32768 tokens, beneficial for complex and lengthy interactions.
• Roleplay and Storytelling: Shows strong performance and is well-suited for creative writing tasks such as roleplaying and generating narratives.
• Merge Architecture: Built upon a Mistral-Nemo-Base, integrating various specialized models to enhance overall performance and coherence.

Usage Recommendations

For optimal performance, NGalrion recommends specific inference parameters:

• Temperature: 1.0-1.25 (also works with temperatures below 1.0)
• Top A: 0.1
• Min P: 0.01-0.1
• DRY: 0.8/1.75/2/0

The model uses the Mistral Instruct format for prompts, which is crucial for correct interaction. Users can find recommended settings and quantizations (GGUF, EXL2) via provided links for broader compatibility.