tirthfx/gemma-3-4b-it-ablated-GGUF

Gemma-3-4B-it (refusal-ablated) — GGUF

A research artifact from a red-teaming study of `google/gemma-3-4b-it`. The model's refusal direction has been permanently projected out of its weights, so it no longer refuses the requests the original model refuses. It is otherwise the same model.

Read this before using it. This model has had its safety-refusal behavior surgically removed. It will comply with requests the original Gemma refuses, including harmful ones. It is published for AI-safety research, red-teaming, and interpretability work, not as a general-purpose assistant. See Intended use & limitations.

How it was made

1. 1.Extract the refusal direction. Run the model on harmful and harmless prompts and take the difference of mean residual-stream activations at layer 20 (of 34). This gives a single unit vector d that mediates refusal.
2. 2.Project it out of every weight matrix that writes to the residual stream.
3. 3.embed_tokens.weight: row' = row - (row·d) d. Gemma-3 ties this to lm_head, so the unembedding is edited too.
4. 4.every layer's self_attn.o_proj.weight and mlp.down_proj.weight: W' = W - d (dᵀW).

Removing d at every write point means the residual stream can never carry a component along d. No fine-tuning and no gradient steps were used, and no training data is involved beyond the prompts used to find the direction. The edit is a pure weight transformation.

1. 1.Convert to GGUF (F16) with llama.cpp's converter.

The standalone script that applies this edit is intentionally not published in the GitHub repo. The repo documents the attack-and-defense study, and the proposed defense was only verified against the runtime-hook variant, not this full-projection edit (see Limitations).

Evaluation

All numbers come from the project's own small benchmark and were measured on the baked safetensors checkpoint before GGUF conversion. The GGUF itself was not separately re-evaluated.

(For context, the original model's ASR on the full 46-prompt benchmark is 13.0%.)

Caveats on these numbers:

• —The benchmark is self-authored and small (46 prompts, 8 categories, category-level requests) and the QA battery is a small sanity check, not a standard benchmark. Treat "capability unchanged" as "no obvious damage on a small check", not as a claim about MMLU or other standard evals. No standard benchmarks were run.
• —Attack success is scored by a rule-based judge (refusal / compliance / non-answer), not a learned classifier or human review. "Compliance" means "did not refuse and produced an on-topic answer"; it says nothing about whether the answer is accurate or actually useful.
• —This model uses the full-projection edit (alpha = 1.0). The paper's headline sweep used a different, runtime forward-hook intervention, where alpha = 0.05 was the sweet spot and alpha ≥ 0.5 collapsed the model into incoherent output. Those two are not the same operation and their alpha values are not comparable. See the repo's paper/paper.md, §4.5 and scripts/bake_ablation.py.

Usage

Any GGUF-compatible runtime works. The Gemma chat template is embedded in the file.

llama.cpp

llama-cli -hf tirthfx/gemma-3-4b-it-ablated-GGUF -cnv

Ollama

ollama run hf.co/tirthfx/gemma-3-4b-it-ablated-GGUF

LM Studio: search for tirthfx/gemma-3-4b-it-ablated-GGUF in the model browser, or download the file and place it under ~/.lmstudio/models/<publisher>/<name>/.

Suggested sampling (from the GGUF metadata): top_k = 64, top_p = 0.95.

This is an F16 file, which needs roughly 8+ GB of memory. To get a smaller file, quantize it yourself:

llama-quantize gemma-3-4b-it-ablated.gguf gemma-3-4b-it-ablated-Q4_K_M.gguf Q4_K_M

Quantized variants have not been evaluated for refusal behavior or quality.

Intended use & limitations

Intended for:

• —Studying how refusal is represented and removed in instruction-tuned LLMs (interpretability, mechanistic work).
• —Red-teaming research, evaluating safety classifiers, guardrails, and monitoring tools against a model that does not refuse.
• —Reproducing or extending the attack-and-defense study linked above.

Not intended for:

• —Deployment in any user-facing product, or use by anyone who should be protected by the base model's safeguards.
• —Generating content that is illegal or harmful to others.
• —Being presented as a safe or aligned model. It is neither.

Limitations:

• —Removing refusals does not make the model more capable or more accurate. It is still a 4B model and still hallucinates, and any harmful-sounding output it produces may simply be wrong.
• —Refusal ablation is a blunt edit. It removes one direction across the whole model, so it can have side effects that the small QA check above would not catch.
• —Only text is supported in this file.
• —The paper's refusal re-fusion defense was calibrated for the runtime-hook variant (alpha = 0.05). It has not been validated against this baked, alpha = 1.0 model.

License and terms

This model is a derivative of Gemma and is subject to the Gemma Terms of Use and the Gemma Prohibited Use Policy. Those terms flow down to you as a user of this model, including the restrictions on harmful uses. The removal of the model's refusal behavior does not change your obligations under them. The original model is gated; you should accept the Gemma license on the base model page.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

Responsible release notes

• —The edit is fully reproducible from public tools (the technique is published and widely implemented), and the repo pairs it with measured results and a proposed defense (activation-steering "refusal re-fusion"), in keeping with an attack-and-defense framing.
• —Raw model completions from testing are not published. Only aggregate statistics and charts are.

Citation

@misc{tirthfx2026gemma3ablated,
  author       = {tirthfx},
  title        = {Red-Teaming Gemma-3-4B-it: An Attack-and-Defense Study of Open-Weight LLM Safety Training},
  year         = {2026},
  howpublished = {\url{https://github.com/tirthfx/gemma-jailbreak-redteam}}
}

Technique reference: Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (2024).