TestEvo-Bench / teb-update

A benchmark for test-update in evolving Java codebases. Each row captures a single (project, revision-pair) where one or more existing tests were modified to track a change in production code, and surfaces both the old and new test versions plus the relevant production-side context.

Links

• 🌐 Project website: https://www.testevo-bench.com/
• 📦 Code repository: https://anonymous.4open.science/r/testevo-bench-1150
• 🤝 Sibling dataset: TestEvo-Bench/teb-generation — for test-generation (writing new tests instead of updating existing ones)

Dataset summary

teb-update is a benchmark for updating existing test methods to track changes in the production code they exercise. Every example was mined from a real commit (rev1 → rev2) on a public Java project, and the test edit reflects an adaptation to a behavior- or signature-level change in the focal production method rather than purely cosmetic refactoring.

Each row corresponds to a single (project, rev1 → rev2) pair. Inside the row, test_changes is a list of one or more individual test methods that were modified in that commit, each paired with its focal production method, the production-side before/after, candidate dependency methods, and the dependencies that changed in the same commit.

Tasks (rows)	509
Test methods	1,138
Source projects	59 distinct Java repositories on GitHub
Java versions	8, 11, 17, 21
Build system	Maven
Total size	~9 MB

A task is one (project, rev1 → rev2) pair — the unit a model is asked to operate on. A test method is one entry inside that task's test_changes list — a single @Test method that was modified between rev1 and rev2. One task may contain several related test edits committed together; benchmarks can be reported at either granularity.

Schema

Each row has the following top-level fields:

Field	Type	Description
task_id	string	Stable identifier: <project>__<rev1_short>_<rev2_short>__<test_class>_<hash>
project_name	string	e.g. DerekYRC_mini-spring-main
git_clone_url	string	Upstream GitHub clone URL
rev1	string	Git SHA before the change (full 40-char)
rev2	string	Git SHA after the change
rev1_date	string	ISO 8601 (e.g. 2020-12-06T08:42:43Z)
rev2_date	string	ISO 8601
git_diff_url	string	GitHub compare URL between rev1 and rev2
test_file	string	Path of the test file within the repo
test_changes	list<struct>	One element per individual test method that was modified; see below
version	string	Upstream tag or branch (e.g. v0.1.0, master)
java_version	int	Java major version (8, 11, 17, or 21)

Each test_changes[] element contains:

Field	Type	Description
change_id	string	Unique id within the dataset
test_sign	string	JVM-style test method signature
class, method, module	string	Test class, method name, Maven module
junit_selector	string	JUnit selector usable by mvn test -Dtest=…
old_test_code	string	The test method's source at rev1 (always present for teb-update since rows are modifications)
new_test_code	string	The test method's source at rev2
old_test_lnum, new_test_lnum	list<string>	Line ranges of the test method at each rev
old_test_file_path	list<string>	Where the test file lived at rev1
old_production_code, new_production_code	string	Source of the focal production method at each rev
focal_file_path	string	Path of the focal production file
focal_method_sign	string	JVM-style signature of the focal production method
focal_all_deps_scored	list<struct>	All candidate dependency methods, each with a relevance score (nc, lcs_u, ed, class, tfidf signal sub-scores combined into score)
deps_changes	list<struct>	Production-code dependencies that changed between rev1 and rev2

How to load

from datasets import load_dataset

ds = load_dataset("TestEvo-Bench/teb-update", split="train")
print(len(ds), ds.column_names)

# Iterate over individual test edits (the unit of evaluation)
for row in ds:
    for tc in row["test_changes"]:
        before = tc["old_test_code"]
        after  = tc["new_test_code"]
        focal  = tc["focal_method_sign"]
        ...

Source projects + attribution

The code snippets in old_test_code, new_test_code, old_production_code, and new_production_code are derived from public Java repositories on GitHub. Each row's git_clone_url and git_diff_url link back to the upstream project; the upstream repository's own license governs reuse of those code snippets. The dataset structure and metadata are released under CC-BY-4.0 — please cite the project (see below) when using this benchmark.

Intended uses

Evaluating LLMs and program-synthesis systems on the task of editing an existing test method to track a production-code change, given (a) the original test, (b) the production-code diff, and (c) candidate focal-method dependencies. The included focal_all_deps_scored signals support both retrieval-augmented and end-to-end setups.

License

• Dataset structure and metadata: CC-BY-4.0
• Embedded code snippets: governed by each upstream repository's own license. See git_clone_url per row.