README.md

---
license: bsl-1.0
pretty_name: Widget2Code Data V4.1
task_categories:
  - image-to-text
tags:
  - screenshot-to-code
  - react
  - jsx
  - multimodal
---

# Widget2Code Data V4.1

Widget2Code Data V4.1 contains UI screenshots for training and evaluation, plus
renderable JSX reference examples. Prompt-ready image information and reusable
ground-truth evaluation features are stored directly in each screenshot's
`metadata.json`.

V4.1 is a backward-compatible label-quality patch. It keeps the `train/`, `test/`, and
`sft-v4/` paths stable, while repairing six train references that omitted or miscolored a major
background. Git tag `v4.0.0` preserves the pre-patch dataset and `v4.1.0` identifies this release;
consumers should pin one of those tags or an exact commit rather than a moving branch.

## Dataset layout

| directory | samples | contents |
| --- | ---: | --- |
| `train/` | 1,822 | training screenshots and canonical metadata |
| `test/` | 1,000 | evaluation screenshots and canonical metadata |
| `sft-v4/train/` | 1,822 | JSX references and their renders, linked to `train/` |
| `sft-v4/test/` | 1,000 | JSX references and their renders, linked to `test/` |

The canonical screenshot pools use this two-file layout:

```text
train/image_0004/
test/image_0001/
├── image.png
└── metadata.json
```

The SFT reference pools do not duplicate the target screenshots:

```text
sft-v4/train/image_0004/
sft-v4/test/image_0001/
├── code.jsx
├── rendered.png
└── metadata.json
```

An SFT sample's `metadata.json` contains `source_split: train` or
`source_split: test`. Combine that split with the same sample ID to locate its
target screenshot. `rendered.png` is the output of `code.jsx`; it is not the
target image. The test references are generated examples, not ground-truth
evaluation labels, and normal evaluation reads only `test/`.

## Screenshot metadata

Each `train/` and `test/` metadata file has this shape:

```jsonc
{
  "id": "image_0004",
  "split": "train",
  "sha256": "...",          // SHA-256 of image.png
  "category": "tools",     // null when not labelled
  "has_chart": null,        // null when not labelled
  "eval": {
    "layout": { "margin": [], "mask_empty": false, "bbox_ar": 0.0, "area_ratio": 0.0, "n_comp": 0 },
    "legibility": { "text": "...", "contrast": 0.0, "contrast_local": 0.0 },
    "style": { "hue_hist": [], "sat_hist": [], "polarity": [] },
    "fill": { "black": {}, "white": {} }
  },
  "side_info": {
    "dims": [976, 668],
    "ocr": "- `\"Pesquisar\"` at (16.5%, 10.0%) of widget, font-height ≈ 7.2% of widget height",
    "palette": "Target widget palette (top-4, after AA-fringe consolidation):\n  #353A38    64.63%\n  ..."
  }
}
```

`side_info` is the complete prompt-ready representation derived from
`image.png`: exact `[width, height]`, formatted OCR output, and the formatted
dominant palette. Consumers should read these values directly; there are no
`dims.txt`, `ocr.txt`, or `palette.txt` sidecars.

`eval` caches the features that depend only on the ground-truth screenshot:
geometry/layout, OCR text and contrast, color/style, and missing-prediction fill
scores. A benchmark run can reuse these values for every prediction instead of
recomputing the ground-truth side. Pairwise metrics such as SSIM and LPIPS still
require the prediction image.

## SFT metadata

Each `sft-v4` metadata file identifies the canonical target and records
generation provenance:

```jsonc
{
  "source_split": "train",
  "generation": {
    "model": "gemini-3.1-pro-preview",
    "variant": "policy-repaired",
    "code_sha256": "...",
    "render_ok": true,
    "rendered_sha256": "..."
  }
}
```

Additional prompt, renderer, and validation fields are retained where they were
recorded. The six V4.1 repairs use `variant: policy-repaired-v4.1` and add
`repair_revision: v4.1.0` plus `repair_reason: restore-major-background`. All 2,822 references
have a successful render.

Every `rendered.png` is the output of `houstonzhang/w2c-render:1.3.1` under the
`m2` contract, and `rendered_sha256` is its checksum. That renderer draws the
same source the same way twice, which earlier ones did not: Chromium reused a
tile from a previous frame, so about one chart render in two hundred came back
a different picture, and a chart that did not disable its own animation was
screenshotted at whichever frame the clock allowed. Re-rendering the references
is therefore a check, not a coin flip:

```bash
W2C_RENDER_MODE=m2 docker/run.sh 16          # from the widget2code-render checkout
python scripts/verify_references.py --split sft-v4/train
```

## Rebuilding prompt side information

The canonical metadata was prepared with the Widget2Code benchmark 1.2.0 CPU
container. On the 96-vCPU reference host, 64 workers was the measured best
setting:

```bash
for split in train test; do
  docker run --rm --user "$(id -u):$(id -g)" -e HOME=/tmp \
    -v "$PWD/$split:/data" \
    houstonzhang/w2c-bench:1.2.0 widget2code-sideinfo \
    /data --out /data --workers 64 --force
done
```

CPU output is canonical. GPU OCR follows a different numeric path and is not
guaranteed to produce byte-identical OCR text.

## Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Djanghao/Widget2Code-Data-V4",
    repo_type="dataset",
    revision="v4.1.0",
    local_dir="Widget2Code-Data-V4",
)
```

Files are stored raw and per sample. Training code can assemble its preferred
record format after download.