This is an uncensored version of google/gemma-3-4b-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 saw some garbled text from time to time (e.g., "It' my" instead of "It's my").

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

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 most layers (layer 7 to 29), independently. This is combined with a refusal weight that follows a symmetric pattern from 0.05 to a peak of 0.55.

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

TBD.