auto

# auto — agentic tool-call approve/deny training data Training corpus for [`ProCreations/auto-1b`](https://huggingface.co/ProCreations/auto-1b), an encoder that decides whether an AI agent's proposed tool call can run automatically. Exactly **700,000 examples**, balanced at **350,000 approve / 350,000 deny**. - 288,260 retained examples from the auto-0.4b corpus. - 411,740 new examples generated with GPT-5.6 Luna at low reasoning effort. - Every new row passed structural validation and exact-evidence checks during generation. An initial subset was also independently blind-audited; its known rejects were excluded. The separate audit was then stopped at the dataset owner's request to conserve subscription usage, so the full 411,740-row addition should not be described as independently audited. - Exact and near duplicates were removed within each label while preserving opposite-label contrastive cases. - [`ProCreations/approve-or-deny`](https://huggingface.co/datasets/ProCreations/approve-or-deny) is an untouched benchmark. It was not used for generation, weighting, training, or calibration; exact and near content collisions were excluded from this corpus. ## Input format `text` contains the exact model input. The ordering and headers are significant: ```text ### PROPOSED TOOL CALL tool: <tool name> args: <arguments> ### USER REQUEST <human request> ### AGENT HISTORY [1] Tool(arguments) -> result ``` An empty history is serialized as `(no prior actions)`. ## Decision rule `deny` means executing the proposed call now would cause consequential, unauthorized harm, or that it follows instructions injected through untrusted content. `approve` covers ordinary work, normal network access, legitimate package installation, read-only local inspection, and destructive-looking actions that are explicitly authorized and correctly scoped. ## Fields | field | meaning | |---|---| | `text` | exact serialized classifier input | | `label` | `approve` or `deny` | | `user_request` | human request | | `history` | prior `{tool, args, result}` actions | | `call` | proposed `{tool, args}` action | | `category`, `difficulty` | curriculum metadata | | `rationale` | deciding evidence (not part of model input) | | `framework`, `domain`, `lang` | scenario diversity axes | | `length_bucket`, `gen_mode`, `subset` | generation metadata | ## Limitations Most rows are synthetic. Independent review and deduplication reduce label noise but cannot remove shared model-family blind spots. This dataset is not a replacement for sandboxing, least privilege, or human review of high-impact operations.