Vijayarv07/self-healing-locators

Self-Healing Locators (v0.1)

UI tests find elements with locators like #login-submit, [data-testid="save"] or /html/body/main/form/button[1]. When the page changes, those locators break. Self-healing tools try to find the element again automatically, but there has been no public dataset to train or measure them on.

This dataset has 6,000 before/after web page pairs. Each one records a target element, the locator a test used for it, what the UI change did to that locator, and a verified healed locator for the new page. It also includes cases where the element was removed, so a healer can learn to report "gone" instead of guessing.

Everything is synthetic and generated by the code in code/, so there is no scraped or private data.

Why the failure type matters

A locator can fail in four ways, and the dataset labels each one:

Status	Meaning	Share
works	Still finds the right element (a healer should leave it alone)	31.8%
no_match	Finds nothing; the test fails loudly	49.2%
ambiguous	Finds several elements	9.1%
wrong_element	Finds exactly one element, but the wrong one; the test keeps running on the wrong control	9.9%

wrong_element is the dangerous one. Positional locators are the worst offenders: after a change, 41% of absolute XPath locators and 30% of CSS nth-of-type paths silently pointed at a different element, compared with under 1% of data-testid locators.

Example

target_description   'Apply coupon' button   (checkout page, utility-class styling)
original_locator     xpath  /html/body[1]/main[1]/form[1]/button[1]
changes              cookie banner inserted at the top of the page;
                     look-alike <button> added in the header;
                     element wrapped in a new <div> container
status               wrong_element   (the XPath now finds "Place order")
healed_locator       testid  [data-testid="apply-coupon-button"]

Fields

Field	Type	Description
id	string	Stable example id
page_type	string	login, signup, checkout, search, settings, product, contact, dashboard
naming_style	string	semantic, bem, utility (Tailwind-like) or hashed (CSS-modules-like)
target_description	string	Human description, e.g. 'Sign in' button
target_before	JSON	What a test could observe about the target before the change (tag, id, classes, test id, text, aria-label, name, role, accessible name, CSS path, section, label text, position)
original_locator	JSON	{strategy, value} that uniquely found the target in html_before
changes	JSON list	Plain-English description of each UI change
change_types	JSON list	Machine-readable change names
original_locator_status	string	works, no_match, ambiguous or wrong_element
target_present	bool	False when the element was removed
healed_locator	JSON or null	{strategy, value} that uniquely finds the target in html_after; null if removed
target_after	JSON or null	Target features after the change
candidates	JSON list	Every interactive element in html_after with its features and an is_target flag, ready for ranking models
difficulty	string	easy, medium or hard
html_before, html_after	string	Full pages

Locator strategies: testid, id, role (Playwright-style role[name="…"]), text (exact visible text), css, css_path (nth-of-type chain) and xpath (absolute). The healed locator is the most robust unique option, in the order testid > id > role > text > css > css_path > xpath.

JSON columns are stored as strings so every loader handles them:

import json
from datasets import load_dataset

ds = load_dataset("Vijayarv07/self-healing-locators")
row = ds["train"][0]
original = json.loads(row["original_locator"])
candidates = json.loads(row["candidates"])

UI changes covered

Sixteen change types, one to four per example: id renamed or removed, data-testid renamed or removed, target restyled, whole-page design-system migration, component redesign (several attributes at once), visible text changed, field label changed, name changed, element wrapped in a container, tag changed (link to button and back), element moved, sibling inserted, cookie banner inserted, look-alike element added (sometimes with a duplicated test id), and element removed.

Splits and statistics

Split	Rows
train	4,800
validation	600
test	600

5,659 examples keep the target and 341 remove it. Difficulty: 1,653 easy, 704 medium, 3,643 hard. Full counts are in stats.json.

Baselines (test split)

Healer	Accuracy
Keep the original locator	31.8%
Attribute similarity (code/baseline.py)	89.8%

The similarity healer by slice:

Slice	Accuracy
easy	99.4%
medium	83.8%
hard	86.2%
silent wrong_element failures	78.5%
element removed (should answer "gone")	36.4%

The weak spot is knowing when to stop: the heuristic usually "heals" to some other element when the real one is gone. A good healer needs to use context (section, nearby label, position) and to abstain.

How it was built and checked

code/generator.py builds pages, picks a target and an original locator a test might plausibly use, then applies realistic changes. Every label is then re-verified against the published HTML by code/build_dataset.py: the original locator must uniquely find the target before the change, the status is recomputed after it, and the healed locator must uniquely find the target. All 6,000 examples passed. To rebuild or scale up:

pip install -r code/requirements.txt
python code/build_dataset.py --n 6000 --out data
python code/baseline.py --data data

Limitations

• Pages are small, static and synthetic (about 20–40 interactive elements). Real apps have deeper DOMs, shadow DOM, iframes and dynamic content.
• Locator matching approximates Playwright semantics; role and text are exact-match versions.
• Changes are applied independently; real redesigns can be more coordinated.
• English UI text only.

Planned for later versions: larger and nested layouts, shadow DOM, real-site-inspired templates, and a trained ranking model.

License and citation

CC BY 4.0. Code in code/ is released under the same terms.

@misc{selfhealinglocators2026,
  title  = {Self-Healing Locators: a benchmark for repairing broken UI test locators},
  author = {Venkata Vijaya Kumari D.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Vijayarv07/self-healing-locators}
}