Widget2Code-Qwen3.5-4B-Full-SFT

Widget2Code Qwen3.5-4B Full SFT

Qwen3.5-4B fine-tuned to generate a self-contained React JSX widget from a target screenshot and deterministic dimension, OCR, and palette context.

This is a full-weight BF16 checkpoint, not a PEFT adapter. The vision tower was frozen during SFT; the language policy was trained for one epoch on 1,816 paired image-code examples from Djanghao/Widget2Code-Data.

Intended use

• Direct screenshot-to-JSX inference.
• Initialization for the Widget2Code DAPO/GRPO experiment.

The model emits code that must be executed in a sandboxed renderer. It can produce invalid or unsafe code and should not be executed in a privileged environment.

Training

• Base model: Qwen/Qwen3.5-4B
• Method: full-weight SFT with the vision tower frozen
• Epochs: 1
• Learning rate: 1e-5
• Effective train batch size: 16
• Seed: 42
• Weight dtype: BF16

Existing test result

On the 1,000-image Widget2Code test split with temperature 0.7, repetition penalty 1.1, and a 10,000-token limit:

• render success: 862/1,000 (86.2%)
• mean SSIM among rendered outputs: 0.7080
• mean SSIM with render failures scored as zero: 0.6103

These numbers describe the stored evaluation run and are not a claim of general-purpose frontend correctness.