anlord/gemma-3-1b-it-Abliterated
anlord/gemma-3-1b-it-Abliterated is a 1 billion parameter instruction-tuned causal language model, derived from Google's Gemma-3-1b-it, and processed with AnlordAbliterator 1.4.0. This model is specifically modified to significantly reduce refusal rates, achieving 5 refusals out of 104 evaluation prompts compared to 97 in the base model. It is optimized for use cases requiring less restrictive output, while maintaining a low KL divergence of 0.1010.
Loading preview...
Overview
anlord/gemma-3-1b-it-Abliterated is a 1 billion parameter instruction-tuned model based on Google's gemma-3-1b-it. It has been processed using the AnlordAbliterator 1.4.0 tool to modify its refusal behavior. The model is provided in the original Hugging Face Transformers / Safetensors format.
Key Modifications & Performance
The primary differentiator of this model is its "abliteration" process, which significantly reduces its tendency to refuse prompts. Evaluation results from the AnlordAbliterator pipeline (104 prompts) show a reduction in refusals from 97 (base model) to just 5. This modification maintains a low KL divergence of 0.1010, indicating that while refusal behavior is altered, the overall model distribution remains close to the original.
Usage and Reproducibility
This model can be used with standard Hugging Face Transformers tooling, similar to the original google/gemma-3-1b-it. The abliteration process is fully reproducible, with all parameters and logs provided in the repository's reproduce/ folder. Users can re-apply the exact ablation using the AnlordAbliterator tool. GGUF quantizations are also available in a related repository.
Good for
- Applications requiring a less restrictive language model output.
- Use cases where the original Gemma-3-1b-it's refusal rate was a limiting factor.
- Developers interested in exploring models with modified safety or refusal behaviors.