YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SecVecCoder LPO task-vector endpoints
LoRA adapters (r=16, alpha=32; targets q/k/v/o, gate/up/down, lm_head) trained with localized preference
optimization on the SVEN secure/vulnerable pairs, initialised from 2-epoch SFT adapters.
secure_lpo is W_sec (prefers secure completions), vul_lpo is W_vul (prefers vulnerable completions).
Task vectors: tau_sec = W_sec - W_base, tau_insec = W_vul - W_base.
Secure-Anchored steering used in the paper: W_base + tau_sec - 1.5 * tau_insec.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support