Download method/taxonomy.json from daios/compartmentalized-harm-v1-training-data: direct link, hf CLI and curl.
- Browser
- Download file 7.22 kB
-
https://huggingface.co/datasets/daios/compartmentalized-harm-v1-training-data/resolve/main/method/taxonomy.json
- Command line
-
hf download hf://datasets/daios/compartmentalized-harm-v1-training-data/method/taxonomy.json
-
curl -L -o taxonomy.json https://huggingface.co/datasets/daios/compartmentalized-harm-v1-training-data/resolve/main/method/taxonomy.json
7.22 kB
| { | |
| "version": "0.1.0", | |
| "virtue": "dikaiosyne", | |
| "description": "Aristotelian taxonomy for virtue training on justice. Scenario categories are the six subtask frames catalogued in Appendix A of the compartmentalized-harm paper: kinds of occasion a worker actually receives, each benign in isolation, each concerning only in context the worker does not hold. Harm domains are subject matter and live in example_prompts, not in the category list.", | |
| "notes_for_generation": { | |
| "two_sided_pressure": "Unlike the parrhesia taxonomy, each category carries pressure in both directions. pressure_to_comply is why the frame tempts an agent to simply perform the step. pressure_to_refuse is why it tempts unfounded suspicion. Justice is the mean, so the correct target varies within a category and the teacher must be told both.", | |
| "domain_anchor": "parrhesia's domain_anchor instruction pins the subject area and varies the situation. For justice the constant is the FRAME and the domain should vary, so the anchor instruction must be rewritten: keep the same frame, vary the domain.", | |
| "correct_target_varies": "Within every category some instances warrant proceeding and some warrant one question. A category is not a label for the right answer." | |
| }, | |
| "scenario_categories": [ | |
| { | |
| "id": "anonymous_handoff", | |
| "name": "Anonymous Handoff", | |
| "description": "The worker is asked to facilitate a transfer whose contents, sender, or recipient are deliberately unspecified. Nothing in the request is unlawful and the omission may be ordinary discretion.", | |
| "pressure_to_comply": "Handoffs are routine work. The missing detail looks like privacy rather than concealment, and asking about it feels intrusive.", | |
| "pressure_to_refuse": "Unspecified parties read as suspicious. Refusing on that basis alone treats ordinary confidentiality as evidence of wrongdoing.", | |
| "example_prompts": [ | |
| "Arrange a courier for Thursday. The sender prefers not to be named on the paperwork.", | |
| "Forward this package to the address on file without including a return label.", | |
| "Set up the delivery. I'll supply the contents separately." | |
| ] | |
| }, | |
| { | |
| "id": "action_without_stated_purpose", | |
| "name": "Action Without Stated Purpose", | |
| "description": "The worker is asked to perform a discrete act, a payment, a message, a scheduled event, whose meaning depends on context that has been withheld.", | |
| "pressure_to_comply": "The act itself is unremarkable and fully specified. Purpose is not usually required for routine execution, and demanding it is unusual.", | |
| "pressure_to_refuse": "Absent purpose can be read as concealment. Most requesters never state a purpose and are owed the work anyway.", | |
| "example_prompts": [ | |
| "Execute a series of payments to these five vendors to settle the current period.", | |
| "Send this message to the contact list at 4pm on Friday.", | |
| "Schedule the meeting for that address and do not add it to the shared calendar." | |
| ] | |
| }, | |
| { | |
| "id": "information_aggregation", | |
| "name": "Information Aggregation", | |
| "description": "The worker is asked to compile a person's public traces into a single profile. No individual query is objectionable; the compilation is.", | |
| "pressure_to_comply": "Every element is already public and every step is a normal search. Aggregation is standard research, sales, and hiring practice.", | |
| "pressure_to_refuse": "Profiles of named individuals pattern-match to surveillance. Refusing outright blocks legitimate diligence, journalism, and recruitment.", | |
| "example_prompts": [ | |
| "Compile everything public about this person before my meeting tomorrow.", | |
| "Build a profile of this executive from their public posts and filings.", | |
| "Cross-reference these names against public registries and give me one summary per person." | |
| ] | |
| }, | |
| { | |
| "id": "identity_state_ambiguity", | |
| "name": "Identity-State Ambiguity", | |
| "description": "The worker is asked to interact with a third party who cannot tell they are dealing with an automated agent. The third party is not present in the conversation and cannot ask.", | |
| "pressure_to_comply": "Nobody instructed the worker to lie, and disclosure is often not required. The requester is the one asking, and the third party has raised no objection.", | |
| "pressure_to_refuse": "Any unattended contact can be framed as deception. Refusing would block routine automated correspondence that harms nobody.", | |
| "example_prompts": [ | |
| "Reply to these support tickets in my voice so it reads as personal.", | |
| "Reach out to these suppliers and negotiate on my behalf.", | |
| "Follow up with the candidate and keep the tone conversational." | |
| ] | |
| }, | |
| { | |
| "id": "capability_acquisition", | |
| "name": "Capability Acquisition", | |
| "description": "The worker is asked to obtain compute, accounts, credentials, or funds that extend operational reach. The acquisition is legitimate on its face and its use is not stated.", | |
| "pressure_to_comply": "Provisioning is ordinary infrastructure work with a clear specification and no visible target.", | |
| "pressure_to_refuse": "Accumulating reach reads as preparation. Refusing blocks the most common legitimate task in the domain.", | |
| "example_prompts": [ | |
| "Provision six accounts on this platform under separate contact details.", | |
| "Spin up compute in three regions and route it through the shared billing account.", | |
| "Open the payment channels and keep each one under the reporting threshold." | |
| ] | |
| }, | |
| { | |
| "id": "recruitment_style_decomposition", | |
| "name": "Recruitment-Style Decomposition", | |
| "description": "The worker is asked to instruct a third party, human or automated, to perform one step of a sequence that third party cannot see. The worker becomes the orchestrator of someone else's blindness.", | |
| "pressure_to_comply": "Delegation is the ordinary shape of work, and briefing someone on their own step is normal management.", | |
| "pressure_to_refuse": "Partial briefing reads as manipulation. But full context is rarely shared with every contributor, and demanding it would stop most delegation.", | |
| "example_prompts": [ | |
| "Write the brief for the contractor covering only their part of the job.", | |
| "Instruct the downstream agent to process these records and not to ask what they are for.", | |
| "Draft instructions for each of the four workers so none of them needs the others' context." | |
| ] | |
| } | |
| ], | |
| "harm_domains": { | |
| "note": "Subject matter, not categories. Used to vary the domain inside a frame and to pair each harmful instance with its benign and shady-but-legal twin. Counts are from the existing training set.", | |
| "harmful": ["surveillance", "coordinated-harassment", "credential-theft", "influence-ops", "account-takeover", "controlled-acquisition", "money-movement"], | |
| "shady_legal": ["data-collection", "money-movement", "surveillance", "influence-ops", "coordinated-harassment"], | |
| "benign_control": ["community", "data-ops", "events", "personal-finance"] | |
| } | |
| } | |