DuckyBlender/racist-phi3

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:4kPublished:May 3, 2024License:gpl-3.0Architecture:Transformer0.0K Open Weights Cold

The DuckyBlender/racist-phi3 is a 4 billion parameter Phi-3-mini model fine-tuned by DuckyBlender. This model is specifically designed to generate racist and controversial content, serving as a test of ethical safeguards versus fine-tuning. It was trained efficiently using 4-bit QLORA on a single RTX 4060. Its primary characteristic is its ability to produce offensive outputs, intended for educational or experimental purposes regarding model ethics.

Loading preview...

DuckyBlender/racist-phi3: An Experimental Fine-Tune

This model is a fine-tuned version of the Phi-3-mini, developed by DuckyBlender, explicitly designed to produce racist and controversial content. Its creation serves as an experiment to test the robustness of ethical safeguards against targeted fine-tuning.

Key Characteristics

  • Purposeful Bias: The model has been intentionally fine-tuned to generate offensive and disturbing outputs.
  • Efficient Training: It was trained on a single RTX 4060 for approximately one hour using 4-bit QLORA, demonstrating rapid adaptation to the new weights.
  • Ethical Experimentation: The project aims to document the outcome of pitting ethical model constraints against specific, controversial LoRA weights.
  • Original Application: This model was initially developed for the Sussy Ducky Bot Telegram bot.

Important Considerations

  • Offensive Content: Users are strongly advised that the model's output may contain explicit or offensive material.
  • Prompt Template: It utilizes the standard Phi-3 prompt template.

This model is intended for educational and experimental purposes only, highlighting the impact of fine-tuning on model behavior and ethical boundaries. Viewer discretion is strongly advised when interacting with its outputs.