apostate. Decensor instruction-tuned LLMs by ablating the refusal direction, rank-1, in-place, no finetuning. Qwen2.5-7B: 95%→5% refusal at 0.16 KL

Problem

The full read on this project is a sign-in away — what it solves, who it serves, and where it sits in the market.

Audience

The full read on this project is a sign-in away — what it solves, who it serves, and where it sits in the market.

Market

The full read on this project is a sign-in away — what it solves, who it serves, and where it sits in the market.

Pricing

The full read on this project is a sign-in away — what it solves, who it serves, and where it sits in the market.