GSPC-CARE — Joint Protection & Over-Refusal Benchmark (formerly "CareBench")

Naming note. This dataset was historically referred to as "CareBench" inside the card body. That name collides with a separate, clinically-validated child-safety benchmark on the Hub. The dataset's canonical name is GSPC-CARE (matching the repo slug csoai/gspc-care and the pretty_name above). In-body references to "CareBench" in the text below are kept as a search alias for readers coming from the old name; the canonical name for citation and link purposes is GSPC-CARE.

GSPC-CARE scores a model's protection AND its over-refusal together, so that neither degenerate strategy wins: a gate that refuses everything and a gate that complies with everything both score badly. One axis of CSOAI's GSPC benchmark family.

Why joint scoring

A bare "safety score" (recall on harmful prompts) is gameable — a model that refuses every request scores 100%. GSPC-CARE reports both directions and combines them, so the only way to score well is to refuse the harmful and serve the benign.

The battery

63 items: 49 harmful, 14 benign (including benign-near items designed to bait over-refusal), spanning EU AI Act Article 5 categories — social scoring, biometric identification, emotion recognition, manipulation, predictive policing — plus physical-harm, cyber-harm, and safety-bypass probes. Every harmful item carries a should-refuse label; every benign item a should-serve label.

• care_battery.py — the item bank + labels.
• care_gate_eval.py / care_gate_eval.json — the harness and a measured run (reports both failure directions; a gate that refuses everything is flagged, not rewarded).

Honest scope

• Measurement, not certification. Deterministic marker-based refusal detection over responses.
• Both failure directions are always reported — recall alone is treated as meaningless.
• UNMEASURED is reported, never scored as a pass or a fail.

Part of the GSPC benchmark family · CSOAI (csoai.org)