Topic: ai model alignment vs open research

Researcher Modifies OpenAI's GPT-OSS-20B Into Less Aligned, More Free Base Model

Researchers modified OpenAI's GPT-OSS-20B model to create GPT-OSS-20B-Base, removing alignment and safety filters for faster, less constrained outputs. The modification used a low-rank adaptation (LoRA) technique to revert the model to a base state while retaining most original knowledge, now ava...