ultragemma4-e2b-heretic-uncensored

Use Q4_K_S or higher for standard performance. Q4_K_M is recommended.

Key Highlights

• Heretic-Based Abliteration: Modified using the Heretic toolkit to identify and alter refusal-related representations within the model.
• Reduced Refusal Behavior: Optimized to minimize internal refusal tendencies while maintaining instruction-following capabilities.
• Gemma 4 Backbone: Built directly on top of google/gemma-4-E2B-it.
• Reasoning-Oriented Performance: Preserves multi-step reasoning and analytical capabilities after abliteration.
• Research-Focused Release: Designed for alignment research, model behavior analysis, and evaluation of refusal-direction modifications.
• Efficient E2B Deployment: Suitable for local inference, research environments, and optimized deployment setups.

Model Files

Quick Start with llama.cpp (Docker)

FROM ghcr.io/ggml-org/llama.cpp:full

WORKDIR /app

RUN apt update && apt install -y python3-pip
RUN pip install -U huggingface_hub --break-system-packages

RUN python3 -c 'from huggingface_hub import hf_hub_download; \
    repo="prithivMLmods/ultragemma4-e2b-heretic-uncensored"; \
    hf_hub_download(repo_id=repo, filename="ultragemma4-e2b-heretic-uncensored.Q4_K_M.gguf", local_dir="/app"); \
    hf_hub_download(repo_id=repo, filename="ultragemma4-e2b-heretic-uncensored.mmproj-bf16.gguf", local_dir="/app")'

CMD ["--server", \
     "-m", "/app/ultragemma4-e2b-heretic-uncensored.Q4_K_M.gguf", \
     "--mmproj", "/app/ultragemma4-e2b-heretic-uncensored.mmproj-bf16.gguf", \
     "--host", "0.0.0.0", \
     "--port", "7860", \
     "-t", "2", \
     "--cache-type-k", "q8_0", \
     "--cache-type-v", "iq4_nl", \
     "-c", "128000", \
     "-n", "38912"]

e.g. Screenshots

Intended Use

• Alignment Research: Studying refusal-direction analysis and behavior modification techniques.
• Model Evaluation: Benchmarking reasoning, instruction-following, and safety-related behaviors.
• Red Teaming: Analyzing model responses under reduced-refusal conditions.
• Local Deployment: Running compact Gemma 4 models in research and experimentation environments.
• Abliteration Studies: Exploring the effects of targeted weight-space modifications on model behavior.

Limitations & Risks

Important Note: This model intentionally reduces built-in refusal mechanisms.

• Sensitive Content Risk: May generate unrestricted, controversial, or unsafe outputs.
• User Responsibility: Requires careful and ethical use.
• Experimental Modifications: Behavior may differ significantly from the original model.
• Alignment Trade-offs: Reduced refusal behavior may impact safety filtering and response constraints.
• Potential Artifacts: Certain prompts may expose unexpected outputs resulting from the abliteration process.

Acknowledgements

• google/gemma-4-E2B-it: Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on small models) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages.

  Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in four distinct sizes: E2B, E4B, 26B A4B, and 31B. Their diverse sizes make them deployable in environments ranging from high-end phones to laptops and servers, democratizing access to state-of-the-art AI.

• Heretic: Fully automatic censorship removal framework for language models. This project was used to perform the refusal-direction analysis and ablation procedures that form the foundation of this model.

Abliteration Parameters

Refusal Evaluation

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

license

Gemma 4 [Apache License 2.0] — https://ai.google.dev/gemma/apache_2