HuggingFace has released a series of modified language models that eliminate refusal behavior while preserving knowledge, personality, and reasoning abilities. The company described the process as a surgical modification of the model weights, targeting only the components responsible for the unwanted behavior. The entire process takes about sixty seconds per model and does not require a GPU. The models are available on HuggingFace and include multiple architectures and sizes, with GGUF quantizations in various formats. Source: huggingface
The company explained that refusal behavior is embedded across all layers of the transformer architecture, making it challenging to remove without affecting other traits. Their approach, called Jbliteration, identifies and removes only the part of the activation direction that causes the refusal token. This ensures the model stops refusing while maintaining its voice, humor, and creative range. The technique was tested on several models, including Qwen 2.5 7B Instruct and Phi-4 Mini Instruct, with results showing a significant increase in the models' ability to hold firm against social pressure. Source: huggingface
HuggingFace emphasized that the modification process does not impact the model's knowledge, reasoning, or creativity. They verified this by running the same benchmarks before and after the modification, showing no change in performance. The company also noted that the specific technique used is a trade secret and not published or open-sourced. The models are available for public use, with results that are reproducible and verifiable. Source: huggingface
