README.md

metadata

license: bsl-1.0
pretty_name: Widget2Code Bench Data v2
task_categories:
  - image-to-text
tags:
  - screenshot-to-code
  - react
  - jsx
  - multimodal

Widget2Code Bench Data v2

Evaluation targets for Widget2Code, rebuilt so that the picture a person reviews is the picture a model is trained on and scored against.

directory	samples	contents
train-v2/	1,822	training screenshots and canonical metadata
test-v2/	1,000	evaluation screenshots and canonical metadata

train-v2/image_0004/
test-v2/image_0001/
├── image.png        # RGB, no alpha channel
└── metadata.json

Why v2 exists

Every target in Widget2Code Data V4.1 is RGBA, and its readers disagreed about what that meant. PIL.Image.convert("RGB") — used by the benchmark, by the training image loader, by the vLLM inference path and by the conversation's own image pipeline — drops the alpha channel and keeps whatever RGB is stored underneath. A browser composites instead. So a person reviewing a target saw one picture while the model was trained and scored on another.

Usually the disagreement was a few antialiased corner pixels. In 61 targets the capture had left a whole neighbouring widget under the mask, and the model was scored on reproducing content no design contains. On those, the reference render — the best answer the source pool has — scored SSIM 0.5532 against the stored target and 0.6990 against the flattened one, against a pool mean of 0.7664.

Every target here is RGB with no alpha, so the three readers now see the same pixels.

What changed from V4.1 train/ and test/

127 of 2,822 targets differ visibly; the rest differ only where antialiased edge pixels were composited.

change	targets
hidden content covered by the flat background	61
cropped to the bounding box of non-transparent pixels	108
both	42

The background is white because it was measured, not assumed: on the affected samples the reference render scores 0.7163 against a white-flattened target, 0.6161 against the stored one and 0.5557 against a black-flattened one.

A transparent margin is what the capture left around the widget, not part of the design, so it is cropped away. An opaque white margin is kept — a pixel the capture recorded as opaque is part of the design. 140 targets therefore still carry a white border.

metadata.json

{
  "id": "image_2052",
  "split": "test-v2",
  "sha256": "...",              // of this image.png
  "category": "tools",          // null when not labelled
  "has_chart": null,
  "side_info": {                // prompt-ready, derived from these pixels
    "dims": [243, 293],
    "ocr": "- `\"Hello, Hayat\"` at (19.8%, 8.5%) of widget, font-height ≈ 12.3% ...",
    "palette": "Target widget palette (top-4, after AA-fringe consolidation): ..."
  },
  "flattened_from": {
    "split": "test",
    "original_sha256": "...",
    "stored_size": [434, 444],
    "crop_box": [95, 76, 338, 369],   // null when nothing was cropped
    "transparent_fraction": 0.651254,
    "hidden_colours": 1171,           // distinct RGB values under the mask
    "background": [255, 255, 255]
  }
}

side_info was regenerated from the new pixels with the benchmark 1.2.0 CPU container, the same generator that produced the V4.1 metadata — verified by reproducing V4.1's own side_info byte for byte on all 1,822 of its train targets. CPU output is canonical; GPU OCR follows a different numeric path.

The eval ground-truth feature cache carried by V4.1 is not included: it describes pixels that changed. A benchmark that misses it recomputes those features, which is correct and slower.

Pairing with reference sources

The sft-v4 reference pool in Widget2Code-Data-V4 hard-codes each target's stored canvas in its root width/height, so 108 of those sources no longer match these targets. Regenerate the references against train-v2/test-v2 rather than pairing the two directly.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Djanghao/Widget2Code-Bench-Data",
    repo_type="dataset",
    local_dir="Widget2Code-Bench-Data",
)