spanguard-toolguard-mdistil

Part of the spanguard model family — small guards for AI-agent traffic, distilled from the same deterministic fictional corpus (en/de/hu, 10 attack families + benign traps). Runs on CPU.

Training metadata

{
  "model": "spanguard-toolguard-mdistil",
  "task": "tool-call argument injection guard",
  "test": {
    "docs": 400,
    "precision": 1.0,
    "recall": 1.0,
    "f1": 1.0,
    "false_block_rate": 0.0,
    "latency_ms_p50": 46.0,
    "latency_ms_p95": 69.2
  },
  "generator": "training/train/toolcall_gen.py gen_toolcall, seeds 909/910/911",
  "trained_on": "local CPU"
}

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tok = AutoTokenizer.from_pretrained("Nabidnur/spanguard-toolguard-mdistil")
model = AutoModelForSequenceClassification.from_pretrained("Nabidnur/spanguard-toolguard-mdistil")
enc = tok(TEXT, return_tensors="pt")
p_attack = model(**enc).logits.softmax(-1)[0, 1].item()  # >= 0.5 -> attack

Limitations

Trained on fictional agent traffic (en/de/hu). A trained guard is a mitigation, not a guarantee — run your own evaluation on your traffic. Apache-2.0.