DreamFast/gemma-3-12b-it-heretic

DreamFast/gemma-3-12b-it-heretic Overview

This model is an abliterated version of Google's Gemma 3 12B IT, processed using the Heretic tool. The primary goal of this modification is to reduce model refusals and "soft censorship" often present in base models, which can lead to weakened adherence to creative prompts or altered visual outputs in downstream applications.

Key Characteristics

• Base Model: Google's Gemma 3 12B IT, a 12 billion parameter instruction-tuned model.
• Abliteration: Achieved using Heretic v1.1.0, specifically Trial 99.
• Reduced Refusals: The abliterated model exhibits only 7 refusals out of 100 tested prompts, a significant reduction from the original model's 100/100 refusals.
• Minimal Model Damage: A low KL Divergence of 0.0826 indicates that the abliteration process caused minimal damage to the model's overall quality.
• Context Length: Supports a 32768 token context length.

Primary Use Case

This model is particularly suited as an uncensored text encoder for video generation models, such as LTX-2. By removing inherent censorship, it allows for more faithful and direct interpretation of creative prompts, leading to more accurate and less sanitized visual outputs. It is available in various formats, including HuggingFace, ComfyUI (bf16 and fp8), and GGUF quantizations (e.g., Q4_K_M recommended for balanced quality/size) for use with transformers, ComfyUI, and llama.cpp respectively.