sdd-distiller-v1

Semantic DOM Distiller — An ONNX model that scores the importance of each DOM node from 0.0 to 1.0.

GitHub: watilde/sdd | Live Demo

Overview

Modern websites have extremely complex and obfuscated DOM structures. Feeding raw DOM trees into LLMs leads to inflated token costs and degraded reasoning accuracy due to noise.

sdd-distiller-v1 is the core scoring model of the Semantic DOM Distiller (SDD) pipeline — a DOM-to-Specification preprocessing engine optimized for multimodal AI agents such as Amazon Nova Act.

Given a 41-dimensional feature vector representing a DOM node, the model outputs an importance score between 0.0 and 1.0. Nodes below a configurable threshold are pruned, reconstructing a lean "functional DOM tree" that contains only what the AI needs to understand the page.

Model Details

Item	Value
Task	DOM Node Importance Regression (0.0–1.0)
Architecture	GradientBoostingRegressor → ONNX
Input shape	(1, 41) float32
Output shape	(1, 1) float32
Model size	~449 KB
RMSE	0.0347
R²	0.9762
Threshold accuracy	97.01% (t=0.3)
Training samples	50,000 (synthetic)
ONNX opset	17

Input Features (41 dimensions)

#	Feature	Description
0	isHighValueTag	1 if tag is button / input / a / form / nav / h1–h3 etc.
1	isMediumValueTag	1 if tag is h4–h6 / label / section / article etc.
2	isContainerTag	1 if tag is div / span / p / section etc.
3	isInteractive	1 if element is interactable (click, type, select)
4	isClickable	1 if element has click handlers or is a link/button
5	hasTabIndex	1 if tabIndex >= 0
6	hasRole	1 if explicit or implicit ARIA role exists
7	roleBaseScore	Base importance score derived from WAI-ARIA role (0.0–1.0)
8	hasAriaLabel	1 if aria-label attribute is present
9	hasAriaLabelledBy	1 if aria-labelledby is present
10	hasAriaDescribedBy	1 if aria-describedby is present
11	hasAriaRequired	1 if aria-required="true"
12	hasAriaExpanded	1 if aria-expanded is present
13	hasAriaLive	1 if aria-live is present
14	hasTestId	1 if data-testid / data-cy / data-test is present
15	hasText	1 if direct text content exists
16	textLength	Normalized text length (0.0–1.0, capped at 200 chars)
17	isLabelText	1 if text is short and label-like (< 50 chars)
18	isActionText	1 if text contains action words (submit, save, login, 送信, etc.)
19	childCount	Normalized child element count (0.0–1.0, capped at 10)
20	hasChildren	1 if element has child elements
21	isLeaf	1 if element has no children
22	depth	Normalized nesting depth (0.0–1.0, capped at 20)
23	depthPenalty	Multiplier penalizing deeply nested nodes (1.0 → 0.25)
24	fontSizeNorm	Normalized font size (0.0–1.0, capped at 72px)
25	isBold	1 if font-weight >= 600
26	areaRatio	Element area / viewport area (0.0–1.0)
27	isAboveFold	1 if element is within the initial viewport (y < 800px)
28	isLargeElement	1 if width > 200px and height > 30px
29	hasHref	1 if href attribute is present
30	hasAlt	1 if alt attribute is present
31	hasPlaceholder	1 if placeholder attribute is present
32	isRequired	1 if required attribute is present
33	isDisabled	1 if disabled attribute is present
34	inputType	Importance score derived from input[type] (0.0–1.0)
35	headingLevel	Importance derived from heading level h1→1.0, h6→0.17
36	parentIsForm	1 if an ancestor is a <form> element
37	parentIsNav	1 if an ancestor is a <nav> element
38	parentIsTable	1 if an ancestor is a <table> element
39	parentIsInteractive	1 if an ancestor is interactive
40	ancestorScore	Decayed importance score propagated from ancestors

Usage

Node.js (onnxruntime-node)

import * as ort from 'onnxruntime-node';

const session = await ort.InferenceSession.create('./sdd-distiller-v1.onnx');

// Build a 41-dim feature vector (Float32Array) for a DOM node
const featureVector = new Float32Array(41);
// featureVector[0] = isHighValueTag, featureVector[3] = isInteractive, ...

const input = new ort.Tensor('float32', featureVector, [1, 41]);
const result = await session.run({ input });
const score = result.output.data[0]; // 0.0 ~ 1.0

Browser (onnxruntime-web)

import * as ort from 'onnxruntime-web';

ort.env.wasm.wasmPaths = 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.17.3/dist/';
const session = await ort.InferenceSession.create('./sdd-distiller-v1.onnx');

const featureVector = new Float32Array(41);
const input = new ort.Tensor('float32', featureVector, [1, 41]);
const result = await session.run({ input });
const score = result.output.data[0]; // 0.0 ~ 1.0

Python (onnxruntime)

import onnxruntime as rt
import numpy as np

sess = rt.InferenceSession("sdd-distiller-v1.onnx")

# Build a (1, 41) float32 feature matrix
feature_vector = np.zeros((1, 41), dtype=np.float32)
# feature_vector[0, 0] = 1.0  # isHighValueTag
# feature_vector[0, 3] = 1.0  # isInteractive
# ...

score = sess.run(["output"], {"input": feature_vector})[0][0]
print(f"Importance score: {score:.3f}")  # 0.0 ~ 1.0

Training

• Training data: 50,000 synthetic samples generated by using the heuristic scorer as a teacher
• Base learner: sklearn.ensemble.GradientBoostingRegressor (200 estimators, max_depth=5, lr=0.05)
• Export tool: skl2onnx with ONNX opset 17
• Training script: scripts/train_model.py

Top Feature Importances

Feature	Importance
parentIsForm	0.2961
isInteractive	0.1637
isHighValueTag	0.1191
depthPenalty	0.1008
depth	0.0942
roleBaseScore	0.0896
isDisabled	0.0334
parentIsNav	0.0211
hasHref	0.0198
isContainerTag	0.0169

Links

• 💻 GitHub: https://github.com/watilde/sdd
• 🌐 Live Demo: https://watilde.github.io/sdd/
• 📦 npm: semantic-dom-distiller (coming soon)