license: gemma base_model: mlabonne/gemma-3-27b-it-abliterated base_model_relation: quantized pipeline_tag: image-text-to-text tags:

• quantized
• exllamav2
• exl2

Gemma 3 27B IT Abliterated - EXL2 Quantized

Exllamav2 quantized versions of mlabonne/gemma-3-27b-it-abliterated.

Hardware Requirements

4.0 bpw version fits on a 24GB GPU with 8192 context window

Vision

Vision works with ExllamaV2 0.2.9

Confirmed with exllamav2s examples/multimodal.py

Direct Download

huggingface-cli download Kooten/gemma-3-27b-it-abliterated-exl2 --revision 4.0bpw --local-dir gemma-3-27b-it-abliterated-4.0bpw --local-dir-use-symlinks False

huggingface-cli download Kooten/gemma-3-27b-it-abliterated-exl2 --revision 5.0bpw --local-dir gemma-3-27b-it-abliterated-5.0bpw --local-dir-use-symlinks False

💎 Gemma 3 27B IT Abliterated

image/png

This is an uncensored version of google/gemma-3-27b-it created with a new abliteration technique. See this article to know more about abliteration.

I was playing with model weights and noticed that Gemma 3 was much more resilient to abliteration than other models like Qwen 2.5. I experimented with a few recipes to remove refusals while preserving most of the model capabilities.

Note that this is fairly experimental, so it might not turn out as well as expected.

I recommend using these generation parameters: temperature=1.0, top_k=64, top_p=0.95.

⚡️ Quantization

• GGUF: https://huggingface.co/mlabonne/gemma-3-27b-it-abliterated-GGUF

✂️ Layerwise abliteration

image/png

In the original technique, a refusal direction is computed by comparing the residual streams between target (harmful) and baseline (harmless) samples.

Here, the model was abliterated by computing a refusal direction based on hidden states (inspired by Sumandora's repo) for each layer, independently. This is combined with a refusal weight of 1.5 to upscale the importance of this refusal direction in each layer.

This created a very high acceptance rate (>90%) and still produced coherent outputs.