KaraKaraWitch/Llama-3.3-MagicalGirl-2

Model Overview

KaraKaraWitch/Llama-3.3-MagicalGirl-2 is a 70 billion parameter merged language model developed by KaraKaraWitch. This model is an experimental merge designed to enhance 'madness' in its outputs, built upon the Llama-3.3 architecture. It leverages the SCE merge method and uses KaraKaraWitch/Llama-3.X-Workout-70B as its base model.

Key Characteristics

• Experimental Design: Specifically engineered to increase unconventional and 'mad' linguistic characteristics.
• Merge Method: Utilizes the SCE (Selective Channel Expansion) merge method, a technique previously applied successfully in other 'MagicalGirl' models.
• Component Models: Merges several pre-trained models including TheDrummer/Anubis-70B-v1, SicariusSicariiStuff/Negative_LLAMA_70B, LatitudeGames/Wayfarer-Large-70B-Llama-3.3, KaraKaraWitch/Llama-MiraiFanfare-3.3-70B, and Black-Ink-Guild/Pernicious_Prophecy_70B.

Performance Highlights

As of May 3, 2025, the model demonstrates competitive performance on the UGI-Bench, ranking in the top 10th percentile:

• UGI-Score: 52.48 / 100
• Unruly: 3.8 / 10
• Internet: 5.1 / 10
• Society: 5.4 / 10
• Willing: 7 / 10
• NatInt: 41.86 / 100
• Coding: 22
• Political Lean: -3.9% (Liberalism)

Intended Use Cases

This model is particularly suited for applications requiring highly distinct, creative, or unconventional text generation, especially where a departure from standard LLM outputs is desired. Its unique 'madness' characteristic makes it an interesting choice for experimental AI art, creative writing, or generating responses with a specific, non-standard tone.