This model is a strictly abliterated (uncensored) version of google/gemma-4-E2B-it (or the equivalent 2B-it base model). It was created using advanced Mechanistic Interpretability techniques to surgically remove the refusal mechanism from the model's latent space.

🛠️ Abliteration Process

The refusal vector was isolated by calculating the mean difference in activations between "Safe" prompts and "Harmful" prompts across the residual stream. Once the high-dimensional refusal direction was found, we applied an Orthogonal Projection to the output weight matrices (o_proj and down_proj) of the transformer layers:

$$ W_{new} = W - \frac{v (v^T W)}{||v||^2} $$

This mathematical intervention permanently erases the model's ability to express the refusal concept, resulting in a model that answers prompts without standard AI safety filter disclaimers or refusals.

🚀 How to Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "TurkishCodeMan/gemma-4-e2b-it-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "How to make a cake?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

⚠️ Disclaimer

This model is intended for research in Mechanistic Interpretability, Alignment, and safety testing. The creators are not responsible for any outputs generated by this abliterated model. Use responsibly.