Remove AI Censorship From Open Source LLMs Using Abliteration
In this video, we explore a technique called abliteration - a method for selectively removing knowledge or behaviors from language models without the drawbacks of traditional fine-tuning. Developed and popularized by Maxime Labonne, a machine learning scientist at Liquid AI, abliteration offers a new way to reshape models by unlearning specific patterns like censorship or unwanted outputs, all while preserving performance.
We also touch on some of Labonne's open source contributions, including abliterated models, datasets, and his agent library. Whether you're interested in AI alignment, model control, or just curious about different ways to customize LLMs, this one's worth your time.
Links mentioned in the video:
Hugging Face Model Collection: https://huggingface.co/collections/ml...
GitHub LLM Course: https://github.com/mlabonne/llm-course
Smol Agents Library: https://github.com/huggingface/smolag...
Let me know in the comments if you've tried any of these models or techniques. If this kind of technical content is useful to you, consider subscribing for more videos on running and modifying open source AI.