modFDTGPT2-XL is a variant of the original GPT-2 XL model fine-tuned with a corrigibility dataset. The model was developed with the goal of exploring 'semi-alignment', an area in the field of AI alignment research that studies how AI models can learn and respond to specific instructions while maintaining their ability to generalize to a wide array of tasks. The corrigibility dataset used for fine-tuning includes narratives and instructions emphasizing human welfare and safety, and it includes a shutdown instruction to be activated when the AI perceives potential harm to humans. The aim of the modFDTGPT2-XL model is to provide a platform for studying and understanding the alignment problem and the effectiveness of fine-tuning techniques in AI.
