grimjim/gemma-3-12b-it-projection-abliterated
grimjim/gemma-3-12b-it-projection-abliterated is a 12 billion parameter instruction-tuned language model derived from Google's Gemma-3-12b-it. It has undergone "projected abliteration" to reduce refusal rates while maintaining safety awareness. This model is notable for its high compliance rating, achieving a 9.8 WC/10 score on the UGI Leaderboard, making it suitable for applications requiring reduced content refusal.
Loading preview...
Model Overview
This model, grimjim/gemma-3-12b-it-projection-abliterated, is a 12 billion parameter instruction-tuned language model based on google/gemma-3-12b-it. It features a context length of 32768 tokens.
Key Differentiator: Projected Abliteration
The primary distinction of this model is the application of "projected abliteration". This technique was used to modify the model without subsequent fine-tuning to repair any damage. The intended outcome is a model that:
- Refuses less often: It is designed to be less prone to declining user requests.
- Retains safety awareness: Despite reduced refusal, it still maintains an understanding of safety guidelines and potential harms.
Performance and Compliance
This model has demonstrated strong performance in compliance metrics. It recently achieved a 9.8 WC/10 rating on the UGI Leaderboard, tying for first place in compliance. This indicates its effectiveness in adhering to specified guidelines while processing requests.
Use Cases
This model is particularly well-suited for applications where:
- Reduced refusal rates are desired: For tasks where a model that is less likely to decline prompts is beneficial.
- Safety awareness is still critical: When maintaining an understanding of harmful content is important, even with a lower refusal threshold.
- High compliance is a priority: Its strong performance on compliance benchmarks suggests suitability for regulated or sensitive environments.