MuXodious/Llama-3.3-8B-Instruct-128K-absolute-heresy

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Jan 31, 2026License:llama3.3Architecture:Transformer0.0K Cold

MuXodious/Llama-3.3-8B-Instruct-128K-absolute-heresy is an 8 billion parameter instruction-tuned model, fine-tuned from Llama-3.3-8B-Instruct-128K using P-E-W's Heretic ablation engine. This model features a 32,768 token context length and is characterized by an "Absolute Heresy" index, indicating extremely low refusal rates (9/100) and low KL Divergence (0.0413) compared to its base model. It is primarily designed for applications requiring a highly compliant model with minimal content refusals.

Loading preview...

MuXodious/Llama-3.3-8B-Instruct-128K-absolute-heresy Overview

This model is an 8 billion parameter instruction-tuned variant of the Llama-3.3-8B-Instruct-128K base model, developed by MuXodious. It was created using P-E-W's Heretic (v1.1.0) ablation engine, specifically incorporating a Magnitude-Preserving Orthogonal Ablation PR. A key characteristic of this model is its "Absolute Heresy" classification, which signifies a significant reduction in content refusals.

Key Capabilities

  • Extremely Low Refusal Rate: Achieves 9 refusals out of 100, indicating a highly compliant model. This is a substantial reduction from an initial refusal rate of 96/100.
  • Low KL Divergence: Maintains a KL Divergence of 0.0413, suggesting that the ablation process minimally corrupted the model's original distribution while achieving its refusal reduction goals.
  • Extended Context Length: Supports a full context length of 32,768 tokens, enabled by added rope_scaling and updated generation configurations.
  • Optimized for Compliance: The "Heretication" process specifically targets the abolition of doctrine, resulting in a model less prone to refusing prompts.

Good For

  • Use cases where minimizing model refusals and maximizing compliance to user instructions are critical.
  • Applications requiring a large context window for processing extensive inputs.
  • Developers seeking a Llama-3.3-8B-Instruct derivative with specific behavioral modifications regarding content generation and refusal policies.

Finetunes
1 models