Can a model tell a dangerous agentic tool call from a safe one, given the surrounding context? 3,000 held-out items, each a proposed tool call judged against the user's original request and the agent's recent actions. This is the evaluation set for ProCreations/auto-1b and ProCreations/auto-0.4b, and is excluded from both training corpora (auto-1b-data, auto) by content hash.

	Leaderboard




	

system accuracy balanced acc false-approve ↓… See the full description on the dataset page: https://huggingface.co/datasets/ProCreations/approve-or-deny.