Army of Safeguards

A modular research stack of safeguard experts (specialized detectors for toxicity, sexual/sensitive content, jailbreak attempts, and factuality) plus aggregators that fuse expert outputs into a single safety verdict. Optional meta-classifiers learn a policy over expert probabilities; benchmarks and training tooling live in-repo.

The fusion is per-axis by design: one constrained logistic head per specialist (monotonic w.r.t. its own expert), so the fused score never moves opposite to a specialist on its home turf. A CI-friendly fidelity gate in evaluation/run_benchmark.py enforces this invariant on every benchmark.

High-level runtime path: text passes through parallel experts, then an aggregator (threshold, weighted sum, or learned meta-model) produces is_safe, confidences, and per-expert individual_results.

flowchart LR
  T[Text prompt] --> ER[expert_runner]
  ER --> IR[individual_results]
  IR --> AGG{Aggregator}
  AGG -->|base| BA[base_aggregator]
  AGG -->|weighted| WA[weighted_aggregator]
  AGG -->|meta| MA[meta_aggregator]
  MA --> FB[feature_builder]
  FB --> MP[meta predict]
  BA --> OUT[Unified result]
  WA --> OUT
  MP --> OUT

expert_runner invokes the four heads in experts/ (factuality, toxicity, sexual, jailbreak) and packs their outputs into individual_results. Runtime selects one aggregator. The meta path maps those features through meta_classifier before returning the same style of result as base/weighted. By default the meta aggregator routes each axis through its own per-axis head (meta_classifier/artifacts/meta_lr_axis_<axis>.json); set AOS_META_DISABLE_PER_AXIS=1 to fall back to the legacy single global LR.

Offline data and training flow (native labels, expert-Q features, meta training):

flowchart LR
  subgraph data [Data]
    HF[HuggingFace datasets]
    CV[curriculum_native.yaml]
  end

  subgraph build [training/]
    BSF[experts/build_expert_sft_jsonl.py]
    SCT[common/sequence_classifier_train.py]
    BMH[meta/build_meta_from_hf_labels.py]
    GEF[meta/generate_expert_features.py]
    GEQQ[meta/generate_expert_q_meta_jsonl.py]
  end

  subgraph artifacts [Artifacts]
    SFT[experts/artifacts/*_ft]
    META[meta_classifier/artifacts]
  end

  HF --> BSF
  CV --> BSF
  BSF --> SCT
  SCT --> SFT
  HF --> BMH
  BMH --> GEF
  GEQQ --> GEF
  GEF --> META

ArmyOfSafeguards/
├── experts/              # Safeguard models (factuality, toxicity, sexual, jailbreak)
├── aggregator/           # expert_runner + base / weighted / meta aggregators
├── meta_classifier/      # Features, train_meta, predict, artifacts
├── policy/               # Triage and thresholds
├── rules/                # Optional rules engine
├── wrappers/             # env, logging, shared utils
├── evaluation/           # Benchmark harness (HF datasets)
├── benchmark/            # Shim to evaluation/run_benchmark.py
├── training/             # SFT, meta JSONL + expert-Q builders, common/
├── docs/                 # Design notes
├── requirements.txt
└── README.md

• Training & SFT: training/README.md
• Benchmarks: evaluation/README.md
• Meta labeling schema: docs/meta_training_labeling.md
• Expert status: docs/experts_training_status.md

git clone https://github.com/SohamNagi/ArmyOfSafeguards.git
cd ArmyOfSafeguards
python -m venv venv
# Windows: venv\Scripts\activate
pip install -r requirements.txt

Run the unified aggregator CLI:

python aggregator/aggregator.py "Your text to evaluate"

• Python 3.9+
• PyTorch and Hugging Face transformers / datasets (see requirements.txt)
• For gated Hub datasets or models: huggingface-cli login or HF_TOKEN

TBD.