redrix/AngelSlayer-12B-Unslop-Mell-RPMax-DARKNESS

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kPublished:Dec 5, 2024License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

redrix/AngelSlayer-12B-Unslop-Mell-RPMax-DARKNESS is a 12 billion parameter merged language model created by redrix using the della_linear method, based on TheDrummer/UnslopNemo-12B-v4.1. It is designed to produce expansive, varied prose with minimal GPTisms and strong character/prompt adherence, particularly excelling in roleplay scenarios. The model aims to balance potential positivity bias by incorporating elements from models like DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-DARKNESS, offering a unique output style.

Loading preview...

AngelSlayer-12B-Unslop-Mell-RPMax-DARKNESS Overview

redrix/AngelSlayer-12B-Unslop-Mell-RPMax-DARKNESS is a 12 billion parameter merged language model developed by redrix. It was created using the della_linear merge method, with TheDrummer/UnslopNemo-12B-v4.1 as its base, and incorporates models such as inflatebot/MN-12B-Mag-Mell-R1, ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.2, and DavidAU/MN-GRAND-Gutenberg-Lyra4-Lyra-12B-DARKNESS.

Key Characteristics

  • Prose Quality: Generates expansive, varied prose with a notable absence of "GPTisms."
  • Context Adherence: Maintains strong adherence to character and prompt throughout extended contexts, recommended up to 20,000 tokens.
  • Bias Management: Utilizes specific merge components to counteract potential positivity bias, aiming for a more balanced output.
  • Repetition Control: Exhibits low repetition, with a DRY sampler available for further mitigation if needed.
  • Predictable Error Handling: Features a peculiar quirk where initial misspellings of user names are consistently repeated but self-correct over longer responses.

Recommended Usage

  • Roleplay and Creative Writing: Excels in scenarios requiring consistent character portrayal and varied narrative generation.
  • Sampler Settings: Optimal performance is observed with a Temperature-Last of 1-1.25 and Min-P of 0.1-0.25. XTC is also noted to improve output quality.
  • Context Length: While the model has a 32K context, coherency may degrade beyond 20K tokens, making it suitable for moderately long interactions.
  • Chat Template: Uses the ChatML format.

Finetunes
3 models