UNSEEN Gemma 4 Removes AI Refusals With a Single Math Trick
A 26B multimodal Gemma variant with refusal directions surgically removed lands on Hugging Face in GGUF format, ready for llama.cpp.
PRO- New abliterated 26B Gemma 4 multimodal model UNSEEN hits 17K downloads on Hugging Face.
- Ships as GGUF from IQ2_M up to Q8_0 for local inference.
- Runs one-line via llama.cpp, Ollama, LM Studio, or vLLM with vision support.
- Uses abliteration to project the refusal direction out of the weights, no retraining.
- Pitched at bulk image captioning and dataset generation where standard VLMs refuse.
- Underscores that RLHF safety training is a shallow, easily reversible layer.
UNSEEN_Gemma_4_26B_NSFW-GGUF is a community-released model quietly climbing Hugging Face download charts, with roughly 17,000 pulls already logged. Built on Gemma 4's 26B multimodal architecture, it has been abliterated, quantized to GGUF, and packaged for local inference through llama.cpp, Ollama, LM Studio, and vLLM. The uploader's pitch: a locally runnable image-to-text model with the refusal layer removed, aimed at unrestricted captioning and dataset generation.
Abliteration: what it actually does
Abliteration is a post-training method that skips retraining entirely. It uses representation engineering to locate the specific direction in a model's activation space responsible for refusal, then orthogonalizes the model weights against that direction. The 2024 paper Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., NeurIPS 2024) showed that across 13 open chat models ranging from 1.8B to 72B parameters, refusal behavior is governed by roughly a one-dimensional subspace of the residual stream. Remove that direction and the model stops refusing; restore it and the model refuses even harmless requests.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
