--- pretty_name: studysource task_categories: - image-to-image tags: - sketch - svg - drawing - human-study configs: - config_name: default data_files: - split: train path: data/train-*.parquet --- # studysource Source material for a human study of drawing ability. **One row per (task, sample, backbone)**, holding the two renderings a rater compares: the untrained backbone's prediction (`baseline`) and our fine-tuned checkpoint's prediction on the same sample (`ours`). | column | | |---|---| | `task_type` | `sketch_completion` or `draw_object` | | `class` | the object the sample asks for | | `model` | the backbone: `Qwen-VL-3.5 27B`, `Qwen-VL-3 32B` or `Gemma-4-E4B` | | `sample_id` | `:` on sketch completion, the class on Draw Objects — `class` alone is not unique on sketch completion (38 classes x 3 samples) | | `baseline` | PNG, 512x512 — the untrained backbone's prediction, rendered | | `ours` | PNG, 512x512 — our checkpoint's prediction on the same sample | 642 rows: (114 sketch completion + 100 Draw Objects) x 3 backbones. Every PNG is also a plain file in `images////.png`, so it has its own URL; `manifest.json` lists them all with the run each came from. ## The two tasks **Sketch completion** (114 samples) — the model is given a partial sketch as an image and draws the missing strokes. The panel shows the input sketch with the completion over it, in blue. **Draw Objects** (100 samples) — the model is given only the name of an object and draws it from scratch, with no canvas stated. Each drawing is centred on its own bounding box and scaled to the panel, which is what the scorer sees. ## Rendering Both panels of a pair are identical in every respect but the prediction. * Sketch completion: the completion is drawn at **5 canvas units on a 256 canvas** — the same pen the input sketch is drawn with, so a completion is exactly as thick as the strokes it continues. * Draw Objects: the stroke is pinned in source units (3.8 px at the scorer's 224 px input) so it has a constant apparent width whatever the drawing's own extent. * A prediction that parses to no path at all is rendered as the panel without strokes — the bare input sketch, or a blank canvas. 14 of the 1284 panels are such. Built by `scripts/build_studysource_dataset.py`.