Djanghao/Widget2Code-Data

Widget2Code Data

Widget screenshots paired with the evaluation evidence that depends only on them, plus self-contained image-code examples.

Directory layout

train/       # 1,822 widget screenshots
test/        # 1,000 widget screenshots
sft/         # 1,816 complete examples
sft-v3/       # raw Gemini 3.1 Pro generations, split into train/ and test/
sft-v3-fixed/ # policy-repaired, fully renderable versions of sft-v3
agentic/     # 232 teacher/student correction trajectories and their renders

verify_draft_qwen35_4b_20260827/        # full-SFT 4B outputs on train and test
verify_draft_qwen35_9b_lora_20260827/   # merged LoRA-SFT 9B outputs on train and test
verify_draft_qwen35_27b_20260827/       # merged LoRA-SFT 27B outputs on train and test

sft-v3/<train|test>/image_0004/
sft-v3-fixed/<train|test>/image_0004/
├── image.png        # the widget
├── metadata.json    # labels + precomputed evaluation intermediates
├── ocr.txt          # OCR evidence, as fed to a prompt
├── palette.txt      # palette evidence, as fed to a prompt
├── code.jsx         # JSX paired with the target image
├── rendered.png     # sft-v3*: generated code render; absent on raw render failures
├── dims.txt         # exact target width and height
└── evaluation/
    ├── evaluation.json        # metrics for a successful render
    ├── ocr.json               # OCR results for target and render
    ├── evaluation_black.json  # raw render failures only
    └── evaluation_white.json  # raw render failures only

The three verify_draft_* directories preserve one stochastic generation per benchmark sample. Each split has a summary.json; each sample directory has the raw response, extracted widget.jsx, meta.json, and rendered.png when render succeeded. The model sizes used the same inference contract, but their training recipes differ, so these outputs are validation artifacts rather than a pure parameter-scaling ablation.

sft/ is a separate split, not a view of train/: its image.png is the render of code.jsx, not the original screenshot of the same id, and its OCR and palette describe that render.

sft-v3/ and sft-v3-fixed/ use the corresponding original train/ or test/ screenshot as image.png; their OCR and palette files describe that target. rendered.png is the output of code.jsx. The train split contains 1,822 samples (1,706 raw renders and 1,822 fixed renders). The test split contains 1,000 samples (936 raw renders and 1,000 fixed renders). Raw render failures are retained as code examples with the error recorded in metadata.json, and no mismatched render is substituted.

All four sft-v3{,-fixed}/<train|test> splits include prediction-side evaluation caches produced by widget2code-bench-exp 1.0.0. A successful render has evaluation/evaluation.json and evaluation/ocr.json. A raw sample without a render instead has paired black/white fallback evaluations, so every sample is covered without substituting the fixed render. Each split also has evaluation.xlsx and .eval_v1.0.0/metrics/ aggregate summaries.

The train code was generated with the archived v3 prompt. The test code was generated with the v3.1 prompt in core/generation/prompts/widget_simple.md; its prompt version and SHA-256 are recorded in each test metadata.json.

metadata.json

{
  "id": "image_0004", "split": "train",
  "sha256": "…",            // of image.png; a cache is only valid for its bytes
  "size": [976, 668],
  "category": "calendar",   // train/sft: one of 16; test: null (never labelled)
  "has_chart": null,        // test: true/false; train/sft: null (never labelled)
  "eval": {
    "layout":     { "margin": [...], "mask_empty": false, "bbox_ar": ..., "area_ratio": ..., "n_comp": ... },
    "legibility": { "text": "...", "ocr": [[bbox, text, confidence], ...], "contrast": ..., "contrast_local": ... },
    "style":      { "hue_hist": [36], "sat_hist": [30], "polarity": [sign, strength] },
    "fill":       { "black": {...}, "white": {...} }
  }
}

eval holds the half of a benchmark score that depends on the ground truth alone, so an evaluation run reads it instead of recomputing it for every model it scores. Only SSIM and LPIPS against a prediction genuinely need both images. fill is the score of the ground truth against an all-black and an all-white image, used when a prediction is missing.

Values are full-precision floats: float(repr(x)) == x, so the text form round-trips exactly. null means never labelled, not "known to be absent".

Changes from the previous layout

train_cls/ and test_cls/ are gone. They held byte-identical copies of train/ and test/ — 613 MB — to express a 16-way label that is now the category field. charts/ is gone for the same reason: it was 30 test images already present in test/, and has_chart now covers all 1,000.

Files are raw and per sample; training code assembles records after download.