heterodoxin Claim this profile

Name a trait in plain English, and Ethos finds its direction inside the model so you can turn it up, down, or bake it in. No fine-tuning.

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

J space studio (made along with claude)

Graph-guided KV cache compression for memory-efficient LLM inference.

No description.