craigm26/microduck-feedback

Microduck community feedback

What people told Microduck Studio when they corrected a proposed plan or chose between two walking ducks. One JSON object per line, in the duck-feedback/0 format.

• route_correction: a plain-language router proposed labels for one clause (for example "turn left" → turn_left), and the person kept, edited or discarded them. These labels recalibrate and fine-tune the router.
• policy_preference: the person watched two trained walkers do the same thing in simulation and chose one, called it a tie, or said neither walked well. These are pairwise comparisons, with the side each walker was shown on recorded so side bias can be measured.

How records get here

Nobody uploads to this dataset directly, and the app runs no server. A person who sets Share with the community in the app's Settings can press Contribute. That opens a pull request here from their own Hugging Face account. Maintainers check every pull request with duckbatch feedback pull --pr <n> (which fetches it and refuses any record not marked public) before merging it.

• Consent is in every record. Only records whose consent.share is public are accepted. The app sends no others, and the validator refuses them.
• No personal data in the records. A record carries no name, account, device identifier or location. id is a random UUID per record. The only free text is the clause the person typed.
• The pull request shows the contributor's Hugging Face username. That is Hugging Face's record of who opened it; the app tells the person before they send.
• Held out from the start. duckbatch splits by a hash of id: 80% train, 20% held out, deterministic and never re-drawn. Report anything learned on the held-out part.

Layout

contributions/<UTC date>/<random uuid>.jsonl, one file per pull request.

Licence

CC0-1.0. Contributors are told this in the app before they send.