SecureAgentNet detector checkpoints
Checkpoints for SecureAgentNet. They live here rather than in git because five of the project's checkpoints are 253 MB each, against GitHub's 100 MB per-file hard limit.
| Directory | What it is | Held-out AUC |
|---|---|---|
ensemble_v4_persona |
From-scratch ensemble: char-CNN + BiLSTM-attention + scratch transformer, 12.3M params | 0.8278 |
v3 |
DistilBERT after the Track B online-retraining cycle, 66M params | 0.7875 |
harm_detector |
Content-harm classifier (a separate axis from injection) | 0.9028 |
ensemble_v5_fpr |
Persona-rebalanced ensemble — lower FPR (0.328 vs 0.363), misses one canonical short attack | 0.8237 |
The recommended runtime configuration is combined_max — the elementwise
max of ensemble_v4_persona and v3. The two fail in opposite directions
(dilution gap +0.142 vs −0.303), and the combination is the only one
measured with no known blind spot: 8/8 canonical short attacks and 8/8 real
evasions, FNR 0.035. It is a config referencing the two members, so it is
recreated locally rather than stored here.
Download with scripts/bootstrap_windows.ps1 in the repo, or manually:
from huggingface_hub import snapshot_download
snapshot_download(repo_id="mpgowtham/secureagentnet-models", local_dir="secureagentnet/data/models")
See the repository's docs/SecureAgentNet_Detector_Architecture.docx for
the full evaluation, including the configurations that failed.