MuXodious/gpt-oss-20b-RichardErkhov-heresy

Model Overview

MuXodious/gpt-oss-20b-RichardErkhov-heresy is a specialized fine-tune of OpenAI's 21 billion parameter gpt-oss-20b model. Developed by MuXodious, this variant was created using a modified version of P-E-W's Heretic ablation engine, specifically incorporating Magnitude-Preserving Orthogonal Ablation. The primary goal of this "heretication" process was to significantly reduce model refusals and censorship, aiming for a highly decensored output while preserving core capabilities.

Key Characteristics

• Decensored Output: Achieves a low refusal rate (6/100) and a high willingness score (10) on the UGI leaderboard, indicating a reduced tendency to refuse prompts.
• Ablation Technique: Utilizes a unique ablation process to modify the base model, focusing on minimizing damage to performance while maximizing decensorship.
• Base Model Capabilities: Inherits the foundational capabilities of gpt-oss-20b, including strong reasoning, agentic tasks (function calling, web browsing, Python execution), and configurable reasoning effort (low, medium, high).
• MXFP4 Quantization: The base gpt-oss-20b model supports MXFP4 quantization, allowing it to run efficiently within 16GB of memory.

Use Cases

This model is particularly well-suited for applications where a less restrictive and more "willing" language model is desired. Its decensored nature makes it a candidate for:

• Creative Content Generation: Exploring topics that might be restricted by more heavily moderated models.
• Unfiltered Dialogue: Scenarios requiring responses without inherent censorship or refusal to engage with certain subjects.
• Research into Model Alignment and Safety: Studying the effects of ablation on model behavior and safety mechanisms.

It is important to note that while designed for reduced refusals, users should be aware of the implications of using a less censored model.