proc1v/screen2ax-tree-silver-qwen3vl-235b

Screen2AX-Tree Silver — Qwen3-VL-235B

Silver (pseudo-labelled) accessibility trees for the macOS screenshots of the Screen2AX-Tree dataset, generated by Qwen/Qwen3-VL-235B-A22B-Instruct in a zero-shot setting.

🔗 Source dataset: macpaw-research/Screen2AX-Tree (Apache-2.0). Screenshots and ground-truth trees are redistributed from there unchanged; the predicted trees and generation metadata are the only new fields in this dataset.

Each sample contains the original screenshot, the human-curated ground-truth AX tree from Screen2AX-Tree, and the AX tree predicted by Qwen3-VL-235B from the screenshot alone. The dataset is intended for knowledge distillation of smaller vision-language models on the Screen2AX accessibility-tree generation task, and as a reference output for studying the failure modes of large frontier VLMs on macOS GUI understanding.

⚠️ This is a silver dataset. The ax_tree_pred field is a model output, not a verified label. Quality is not human-checked; treat it as a noisy supervision signal. The ax_tree_gt field carries the original Screen2AX-Tree ground truth and is the only field suitable as a clean reference.

Dataset structure

A single split with all 1127 samples from the original Screen2AX-Tree (the upstream 85 / 15 train/val split is not preserved here — define your own if you need one).

Fields

Field	Type	Description
id	string	Sample id (matches Screen2AX-Tree filename, e.g. "0" ⇒ 0.png).
image	image	macOS application screenshot (Retina, 2× logical resolution).
image_width	int	Image width in pixels.
image_height	int	Image height in pixels.
ax_tree_gt	string	Ground-truth linearized AX tree (from Screen2AX-Tree).
ax_tree_pred	string	Qwen3-VL-235B predicted AX tree.
model	string	Generator model id (Qwen/Qwen3-VL-235B-A22B-Instruct).
temperature	float	Sampling temperature used (0.0).
max_tokens	int	Generation cap (8192).
prompt_tokens	int	Prompt token count for this sample.
completion_tokens	int	Generated token count for this sample.
latency_seconds	float	Wall-clock generation latency.
node_count	int	Number of nodes in the predicted tree.
parse_success	bool	Whether ax_tree_pred parsed cleanly into a tree.

AX tree format

Each line is one UI element; 2-space indentation encodes the parent-child hierarchy:

Role(subrole) [x1,y1,x2,y2] key1="value1" key2="value2"

• Role — accessibility role with AX prefix retained (AXButton, AXStaticText, ...).
• (subrole) — optional subrole (standard window, close button, switch, ...).
• [x1,y1,x2,y2] — bounding box, normalized to a 0–1000 range (top-left / bottom-right).
• Attributes are included only when present: name="...", desc="...", value="...".

Example:

AXWindow(standard window) [101,38,1101,1038] name="App"
  AXToolbar [0,0,1000,60]
    AXButton [10,15,40,45] name="Save"

Generation procedure

• Model: Qwen/Qwen3-VL-235B-A22B-Instruct, served via vLLM (OpenAI-compatible API).
• Decoding: temperature=0.0, max_tokens=8192.
• Input: the raw screenshot at native (Retina) resolution.
• Prompting: zero-shot. The full system prompt enumerates the valid AX roles and the required output grammar; the user message is "Generate the complete accessibility tree for this screenshot." Every sample uses an identical prompt — no few-shot examples, no CoT, no tool use.
• Post-processing: Markdown code-fence stripping only. Trees that fail to parse are retained verbatim and flagged via parse_success=false.

Loading

from datasets import load_dataset

ds = load_dataset("proc1v/screen2ax-tree-silver-qwen3vl-235b", split="train")
print(ds)

ex = ds[0]
ex["image"].show()
print(ex["ax_tree_pred"])

Intended uses

• Distillation / pseudo-label SFT of smaller VLMs (e.g. Qwen2-VL-7B, Qwen2.5-VL-7B, ShowUI) on macOS accessibility-tree generation.
• Error analysis of frontier-scale VLMs on hierarchical GUI understanding.
• Benchmark warm-up data when full Screen2AX-Tree is not enough — silver labels cover the same screenshots; the ground-truth labels live in ax_tree_gt.

Out-of-scope uses

• Reporting absolute model performance: ax_tree_pred is a noisy signal; use ax_tree_gt for evaluation.
• Production accessibility tooling: the predicted trees have not been validated by humans and are unsuitable for assistive-technology end users.

Limitations and known issues

• Coordinate scaling. Screenshots are 2× Retina (logical macOS points × 2); the 0–1000 bbox range is normalized to image dimensions. Convert to pixels by multiplying by image_width / 1000 and image_height / 1000.
• Generator quality is mixed. Even a 235B-parameter model under-recalls deep trees — many predictions miss subtrees, hallucinate roles, or flatten hierarchies. Filter on node_count or parse_success if needed.
• Truncation. Some predictions hit max_tokens=8192 and were cut off mid-tree.
• Surrogate codepoints in the original Screen2AX-Tree have been stripped from text values during linearization to keep the dataset Arrow/Parquet-safe.

Source dataset and license

Derived from macpaw-research/Screen2AX-Tree (Apache-2.0). Screenshots and ax_tree_gt are redistributed unchanged under the same license. The ax_tree_pred field and other generation metadata are released under Apache-2.0 by the dataset author.

If you use this dataset, please also acknowledge the original Screen2AX work and the generator model:

• Screen2AX-Tree — macpaw-research/Screen2AX-Tree
• Qwen3-VL — Qwen/Qwen3-VL-235B-A22B-Instruct