darkc0de/Rotor_24B_V.1-heretic

Rotor_24B_V.1-heretic Overview

This model, darkc0de/Rotor_24B_V.1-heretic, is a 24 billion parameter decensored version of OddTheGreat/Rotor_24B_V.1, created using the Heretic v1.2.0 abliteration tool. The primary goal of this modification was to significantly reduce model refusals, achieving 3/100 refusals compared to 98/100 in the original model, as measured by KL divergence of 0.0906.

Key Capabilities

• Decensored Output: Demonstrates a substantial reduction in refusal rates, making it suitable for a wider range of creative and roleplay applications.
• Roleplay Optimization: Specifically designed to replace Circuitry as a main roleplay model, with a focus on generating less "mechanical" and more natural prose.
• Instruction Following: Excels at understanding and adhering to complex instructions.
• Long Context Handling: Capable of managing long dialogues and maintaining coherence over extended conversations, tested with a 12k context length (model supports 32768 tokens).
• Multilingual Support: Tested for assistant and roleplay tasks in Russian and French, showing good performance.
• Balanced Content: Handles both SFW and NSFW content while maintaining character consistency.

Good For

• Creative Writing & Roleplay: Ideal for generating engaging narratives, character interactions, and text adventures where natural prose and instruction adherence are crucial.
• Decensored Applications: Use cases requiring a model with significantly reduced content filtering or refusal behaviors.
• Long-form Dialogue: Scenarios demanding sustained, coherent conversations over many turns.
• Multilingual Interaction: Applications involving roleplay or assistant tasks in Russian and French.