HEART77/VIGIL-Qwen3.5-4B-L3-Replication

VIGIL Qwen3.5-4B L3 Replication LoRA

Cross-scale replication of the VIGIL L3 wrong-image rejection-ranking method. It was trained for one epoch on the same strict 1,000 paired groups and uses a neutral prompt.

Selected-checkpoint results: original accuracy 70.57%, CVE 70.57%, wrong-image LPE 14.89%, OAR 1.06%, fixed-300 agreement 71.33%, and count agreement 64.24%. The ranking mechanism transfers to 4B, but this checkpoint does not dominate the earlier 4B causal adapter because count agreement regresses. It is therefore published as a research replication checkpoint, not the primary 4B deployment model.

This is a PEFT adapter and requires Qwen/Qwen3.5-4B. See the VIGIL GitHub repository for the exact protocol and limitations.