eulogik / suraksha-450m

SURAKSHA 450M: calibrated guard decision model (prompt-injection + MCP tool-risk, Hinglish, ONNX/CPU, Jev-compatible)

450M open-weight guard by Eulogik . One forward pass turns a state plus typed questions into calibrated probabilities: P(prompt-injection), P(jailbreak), P(PII-leak), P(tool-risk). English plus Hinglish plus Hindi. Apache-2.0, self-hosted, $0.

Code and full evals: github.com/eulogik/suraksha . Dataset: eulogik/suraksha-guard-38k . GGUF: eulogik/suraksha-450m-GGUF (74/74 parity, serve with recent llama-server).

What is in this repo

• phase_b.pt : v2 final guard checkpoint (LoRA r16 + concept bottleneck + head, 7000 steps + 50 RLCD on a Mac mini M4).
• suraksha-guard.onnx : CPU artifact, 37/37 parity with torch, 84ms per decision.
• temperatures.json : per-bucket fitted temperatures.
• eval/*.json : raw heldout results. assets/reliability.png : reliability diagram.

Loading needs the suraksha package (custom head): pip install git+https://github.com/eulogik/suraksha , then SurakshaAgent(device="mps", checkpoint_path="phase_b.pt") . See the GitHub README for byte-exact usage, server, scan CLI, and Ollama.

Numbers (measured, frozen heldout n=3058)

Jev figures quoted in docs are third-party published, never measured here. Zero-shot and fine-tuned numbers are never mixed. Trust Card output is a CI-grade credential, not a pentest or legal verdict.

Limits

512-token en context (1024 multi), head-truncate with truncated:true . JSON tool calls only for the risk head. Hinglish trails English. ONNX gives parity, no speedup claimed.

License

Apache-2.0. Base laya (ConvAI Innovations, Apache-2.0). Banking77 rehearsal rows (PolyAI, CC-BY-4.0).

Keywords: prompt injection detector, mcp guard, tool risk classifier, jailbreak detector, pii leak detector, hinglish guardrails, system one model, jev alternative, on-device classifier, apache 2.0 guard.