lordx64 / phantom-kv
git clone https://github.com/lordx64/phantom-kv.git
Trains a KV-cache prefix of learned key/value tensors that suppresses refusal in LLMs, loadable and unloadable without modifying model weights.
Why you might care
For researchers studying refusal behavior who want a reversible, weight-preserving alternative to abliteration. The base model stays byte-identical, and the graft is megabytes rather than a full checkpoint, so it can be swapped mid-session. The README includes demos and technique docs for Qwen-family models.
+0
stars, 24h · #37,467
+0
stars, 7 days · #55,682
+~184
stars, 30 days · #5,600
273
★ total
42
forks · 1 open issues
Star history (as recorded here)
LanguagePython
LicenseMIT
Created2026-09-17 (1mo ago)
Last push1mo ago
Watchers1
OwnerUser
GitHub topics
Links
README
Loading README…
Blurb generated by claude-haiku-5-5 9d ago. Star deltas marked ~ are extrapolated from partial history.