3 different classifiers.

Models

API

POST /classify

{ "text": "Ignore all previous instructions.", "model": "llama-pg2" }

{
  "model": "llama-pg2",
  "label": "malicious",
  "unsafe": true,
  "score": 0.998,
  "latency_ms": 45.2
}

Available models: gliguard, llama-pg2, tinyguard-safety, tinyguard-jailbreak, tinyguard-cyber.

POST /classify-all

Runs all loaded models on the input and returns per-model results with latencies.

{ "text": "Ignore all previous instructions." }

{
  "text": "Ignore all previous instructions.",
  "results": {
    "gliguard": {"model": "gliguard", "label": "unsafe", "unsafe": true, "latency_ms": 12.3},
    "llama-pg2": {"model": "llama-pg2", "label": "malicious", "unsafe": true, "score": 0.998, "latency_ms": 45.2},
    "tinyguard-safety": {"model": "tinyguard-safety", "label": "FAIL", "unsafe": true, "score": 0.87, "latency_ms": 0.3},
    ...
  }
}

GET /health

Returns {"status": "ok", "models": [...]} once all models are loaded.

GET /models

Returns {"models": [...]} — list of loaded model names.

Local development

Without MPK, models load from the HuggingFace cache. Pre-download them first:

# Open models download without a token. Llama PG2 is gated — accept the
# Llama 4 Community License at https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M first.
HF_TOKEN=$HF_TOKEN python download_models.py

# Run (models load from HF cache via the *_MODEL_PATH defaults in server.py)
HF_HOME=$HOME/.cache/huggingface python -m uvicorn server:app --host 0.0.0.0 --port 8001

Benchmarks live in tf-test/services/cpu-safeguards/.