Overview

sh0ck0r/L3.3-Shakudo-70b-heretic is a 70 billion parameter language model with a 32K context length, derived from Steelskull/L3.3-Shakudo-70b and decensored using Heretic v1.2.0. This model is engineered through a sophisticated multi-stage merging process, including a technique called "weight twisting," to enhance its creative writing and roleplaying abilities while significantly reducing refusal rates compared to its original base model.

Key Capabilities & Features

• Decensored Output: Modified to reduce refusal behaviors, offering more open-ended responses.
• Multi-Stage Merge Architecture: Built from a complex fusion of several base models, including a cognitive and tool-use focused foundation (L3.3-Cogmoblated-70B).
• Enhanced Creative Writing: Utilizes SCE merging to develop rich prose and unique narrative "flavor."
• Strong Roleplaying Depth: Incorporates Della_Linear merging to integrate robust roleplaying capabilities.
• Weight Twisting: Employs an ablated base model (nbeerbower/Llama-3.1-Nemotron-lorablated-70B) to pre-align against refusal patterns, influencing the final model's response style.

Performance Highlights

• Refusal Rate: Achieves a significantly lower refusal rate of 12/100 compared to the original model's 64/100.

Recommended Use Cases

• Creative Writing: Ideal for generating imaginative stories, prose, and descriptive text.
• Roleplaying: Excels in interactive narrative and character-driven roleplay scenarios.
• Open-ended Text Generation: Suitable for applications requiring less constrained and more exploratory outputs.

Recommended Sampler Settings

• Static Temperature: 1.0 - 1.2
• Min P: 0.02 - 0.025
• DRY: Multiplier: 0.8, Base: 1.74, Length: 4-6