pi-lossy-oneshot-5000

One-shot SFT dataset for graphic design editing, generated using the Perturb-and-Invert (P&I) method with metadata-grounded perturbation planners.

What's in it

Each example trains a model to output the full inverse trajectory (all tool calls) in a single response given the initial design state and a natural-language edit request. Trajectories are composed from four semantic planners (Style, Readability, Contrast, Palette) applied to cyberagent/crello templates, ensuring every edit is semantically grounded and invertible.

Lossy inverses: palette/saturation inverse steps use negated deltas rather than per-element color snapshots.

Splits & size

Split	Examples
train	4,500
validation	500

Schema

Field	Type	Description
id	string	Unique example ID
source	string	Perturbation class (composite)
lossy	bool	Whether the inverse used lossy mode
messages	list	[system, user, assistant] chat messages

Usage

from datasets import load_dataset
ds = load_dataset("monish-adobe/pi-lossy-oneshot-5000")