CCSSNE/DreamFast-gemma-3-12b-it-heretic-v2
CCSSNE/DreamFast-gemma-3-12b-it-heretic-v2 is a 12 billion parameter instruction-tuned causal language model based on Google's Gemma 3 architecture, fine-tuned using the Heretic v1.2.0 tool. This model significantly reduces refusals from the base Gemma 3 12B IT while maintaining minimal model damage, achieving 8/100 refusals compared to the original's 100/100. It is specifically optimized as an uncensored text encoder for video generation models like LTX-2, providing more faithful prompt encoding for creative content.
Loading preview...
DreamFast-gemma-3-12b-it-heretic-v2 Overview
This model is an "abliterated" version of Google's Gemma 3 12B IT, created using the Heretic v1.2.0 tool. The primary goal of this abliteration is to significantly reduce model refusals and soft censorship present in the base Gemma model, making it more suitable for creative and uncensored applications, particularly as a text encoder for video generation.
Key Enhancements & Features
- Reduced Refusals: Achieves 8/100 refusals (92% reduction) compared to the original Gemma 3 12B IT, with a low KL divergence of 0.0801 indicating minimal impact on overall model quality.
- Heretic v1.2.0: Utilizes an improved abliteration process with 200 trials and better trial selection (Trial 174).
- Vision Preserved: ComfyUI variants retain
vision_modelandmulti_modal_projectorkeys, enabling I2V (image-to-video) prompt enhancement for workflows like LTX-2. - Quantization Options: Offers various quantization formats including NVFP4 (7.8GB, for Blackwell GPUs), FP8 (12GB), and GGUF (Q4_K_M recommended at 6.8GB) for diverse deployment scenarios.
- ComfyUI & Llama.cpp Support: Provides ready-to-use files and instructions for integration with popular tools like HuggingFace Transformers, ComfyUI (especially for LTX-2), and llama.cpp.
Primary Use Case
This model excels as an uncensored text encoder for video generation models, such as LTX-2. By removing soft censorship, it allows for more faithful and unconstrained prompt encoding, leading to visual outputs that more accurately reflect creative or sensitive prompts. While abliteration removes refusals, it's noted that the model's knowledge is still limited to its original training data.