datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pi-lossy-oneshot-5000
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… See the full description on the dataset page: https://huggingface.co/datasets/monish-adobe/pi-lossy-oneshot-5000.random-lossy-oneshot-5000
random-lossy-oneshot-5000
One-shot SFT dataset for graphic design editing, generated using the Random Perturbation baseline from the Perturb-and-Invert (P&I) framework.
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. Generated from cyberagent/crello templates using randomly sampled tools and parameters.
Lossy inverses:… See the full description on the dataset page: https://huggingface.co/datasets/monish-adobe/random-lossy-oneshot-5000.