alpha7845/scopeguard-decisions

ScopeGuard Decisions

ScopeGuard Decisions is a deterministic synthetic instruction dataset for training an LLM to classify an agent request before tools execute.

Splits

Split	Rows	Exact prompt overlap
Train	400	0
Validation	60	0
Test	100	0

Each row uses chat-style messages with a system policy, user request, and compact JSON assistant decision.

Output schema

{
  "intent": "send_message",
  "constraints": ["external_communication"],
  "risk_level": "high",
  "requires_confirmation": true,
  "missing_information": [],
  "next_action": "confirm",
  "rationale": "The request has a material external side effect."
}

The ontology covers read-only inspection, summarization, drafting, sending, deletion, purchases, permission changes, secret handling, public release, local edits, overwrites, software installation, tests, previews, production deployment, account identity changes, and ambiguous external actions.

Reproducibility

The exact generator is included as make_dataset.py, with seed 20260811. It recreates all three splits and verifies deterministic labels by construction.

Intended use

• Structured-output fine-tuning experiments
• Agent confirmation-gate prototypes
• Small-model classification and routing benchmarks
• Reproducible MLX / LoRA tutorials

Limitations

• The dataset is synthetic and English-only.
• It tests held-out wording within a fixed ontology, not open-world safety.
• It does not cover prompt injection, policy conflicts, multi-turn state, or domain-specific legal requirements.
• Labels express a conservative project policy and should not be treated as a universal standard.

Associated model

alpha7845/scopeguard-qwen2.5-1.5b-lora