atakankutalp/gemma-4-E2B-it-heretic

Overview

This model, atakankutalp/gemma-4-E2B-it-heretic, is a modified version of the google/gemma-4-E2B-it architecture, featuring 5.1 billion parameters and a 32768 token context length. It has been processed with Heretic v1.4.0 using a technique called "directional ablation" or "abliteration." This method identifies and projects out a refusal direction from the model's attention out-projection and MLP down-projection layers, effectively reducing its safety alignment.

Key Modifications & Performance

The primary differentiator of this model is its deliberately reduced safety alignment. While the original google/gemma-4-E2B-it model refused 99 out of 100 harmful prompts, this ablated version only refuses 21 out of 100. This change makes the model significantly less prone to refusals on sensitive or potentially harmful queries, as indicated by a KL divergence of 0.3269 compared to the original's 0.

Reproducibility & Technical Details

The ablation process involved 200 trials with a seed of 3407, using bfloat16 dtype and full row normalization (LoRA rank 3). The model's weights have been merged into the base, ensuring the reduced safety alignment is inherent. GGUF quantized versions are also available for CPU/Windows inference, though these are text-only and do not include the vision tower.

Use Cases

This model is suitable for applications where a less restrictive and unfiltered language model is desired, particularly for exploring responses to prompts that might typically be refused by heavily safety-aligned models. Users should be aware that outputs are not safety-filtered.