pi-lossy-singlestep-5000

Single-step 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 predict a single tool call given the current design state and the history of tool calls applied so far. Trajectories are composed from four semantic planners (Style, Readability, Contrast, Palette) applied to cyberagent/crello templates, producing grounded edit queries and guaranteed-invertible trajectories.

Lossy inverses: palette/saturation inverse steps use negated deltas rather than per-element color snapshots, giving ~5–10× smaller trajectories with bounded per-channel drift.

Splits & size

Split	Examples
train	48,919
validation	5,443

Source: 5,000 Crello templates × ~11 steps each.

Schema

Field	Type	Description
id	string	{source_id}_step{i}
source_id	string	Originating template + perturbation ID
source	string	Perturbation class (composite)
lossy	bool	Whether the inverse used lossy mode
step_idx	int	Index of this step in the trajectory
total_steps	int	Total steps in the trajectory
messages	list	[system, user, assistant] chat messages

Usage

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