LoRA adapter (MLX) for NVIDIA-Nemotron-3.5-Bolt-E4B trained to emit SOC alert
triage decisions as JSON over 7 classes: malicious, suspicious, benign,
expected_admin, authorised_testing, misconfiguration, unknown.
Status: TriageKit v1.0 production release. This adapter is the serving artifact for L1 Copilot recommendations (humans decide) and for L2 Scope A deterministic auto-close of allowlisted routine events (model not involved in that path). Model-driven auto-close (L2 Scope B) remains blocked until 5,000 real analyst labels and a 4-week shadow period exist. GGUF exports fail the quantization regression; do not serve them in production.
Scope Status L1 Copilot (recommendations) READY L2 Scope A (deterministic allowlist auto-close) READY L2 Scope B (model-driven auto-close) BLOCKED SOCBench (DeepTempo) tool-using agent: F1 0.667, precision 1.00 (0 false-malicious) on the stratified smoke slice. See the harness repo docs.
Served via MLX on Apple Silicon, Ollama (GGUF, testing), and vLLM/SGLang (PEFT adapter). See
docs/SERVING.mdin the harness. | Seeevaluation/in the GGUF repo for the release gate.⚠️ EXPERIMENTAL / PILOT (v0.1). 30 iterations on 420 synthetic examples. Not a production model. Must be used with an evidence-grounding guardrail and evaluated on real labeled alerts before any operational decision.
This is not an official NVIDIA or Google model. It is not affiliated with, sponsored by, or endorsed by NVIDIA or Google.
| Base | nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 (MLX 4-bit quant) |
| Method | LoRA (MLX), rank 8, 4 layers, scale 20 |
| Data | infosecn1nja/soc-triage-opensec-chat (derived from tegridydev/opensec-triage, CC BY 4.0) |
| Sequence | max 2048, prompt masking, gradient checkpointing |
| Steps | 30 iters, lr 2e-5, batch 1 |
| Loss | train 2.509 → 0.259; valid 2.833 → 0.207 |
| Hardware | Apple M4 24 GB, mlx-lm 0.32.2 |
# download the adapter
hf download infosecn1nja/soc-triage-nemotron-bolt-e4b-lora --local-dir adapter
# serve (thinking MUST be disabled so the model emits JSON directly)
mlx_lm.server \
--model mlx-community/NVIDIA-Nemotron-3.5-Bolt-E4B-OptiQ-4bit \
--adapter-path ./adapter \
--chat-template-args '{"enable_thinking": false}' \
--port 8080
The system prompt must match training (7-class taxonomy + evidence rules).
Reference harness: github.com/infosecn1nja/soc-triage-harness (product: TriageKit).
| Metric | Value |
|---|---|
| Macro-F1 | 0.954 |
| Accuracy | 0.964 |
| Missed malicious (auto-closed) | 0 |
| Auto-close precision | 1.0 |
| Attack-flag false positives | 0 |
| Escalation false positives | 0 |
| ECE | 0.102 (95% CI 0.06–0.15) |
Harness result: READY for shadow mode (L0) — 11 gates PASS and 3 SKIP (statistical guard: ECE needs n>=200, class recall needs >=10 examples; test n=55). These numbers apply to the synthetic golden set only, not to real telemetry.
evidence_refs; a guardrail that verifies evidence IDs is required.adapters.safetensors — LoRA weightsadapter_config.json — adapter configurationtraining_config.yaml, training_results.md — config and resultsLICENSE-OpenMDW-1.1.txt, NOTICE.txt — license and attributiontegridydev/opensec-triage — CC BY 4.0.