Open-weight & owned stack · Training tuning
Phantom-KV Alters Refusals Without Changing Model Weights
The phantom-kv project uses an approximately 18MB learned KV-cache bank to alter refusal behavior on demand.
Read the original at reddit.comOpens the publisher's site in a new tab