Req2Test-Bench

Code | Paper | Leaderboard | Documentation

Dataset Summary

Req2Test-Bench is a benchmark for requirement-to-test generation in real software projects. It collects 5,086 execution-validated instances from 8 real-world projects in 5 programming languages. Its eval split has 2,051 instances.

Each instance is one Scenario of a project's Gherkin requirements. An agent receives the Requirement (a Gherkin Feature with one target Scenario), the production interfaces that the Scenario targets, and a runnable workspace of the project: its source with the target implementation hidden, and an application runtime in which the project's tests can be run. It must write the Scenario's executable Test Implementation: its Step Definitions and the helper code they need.

Req2Test-Bench runs the generated test against the project's real code and asks two questions. Does it pass? And does it fail on every mutant (a copy of the code with one injected fault) that the developers' own test of that Requirement catches? A test that passes and kills all of those mutants is KillComplete.

To load the tasks and the reference tests:

from datasets import load_dataset
eval_rows = load_dataset("code-philia/Req2Test-Bench", split="eval")
gt = load_dataset("code-philia/Req2Test-Bench", "gt_tests", split="full")

Supported Tasks and Leaderboards

The metrics are Pass@1, KillComplete@1, mutant-kill recall, API F1 and coverage Jaccard (docs/metrics.md). The leaderboard is in results/leaderboard.md.

Languages

The requirements are Gherkin Features written in English. The projects are written in Go, Java, PHP, Python and Ruby. Their tests use the BDD frameworks behave (Python), Cucumber (Java and Ruby), Behat (PHP) and godog (Go).

Project (project)	Eval	Full	Language	Framework	Upstream repository	Upstream licence
Algorand Java SDK (algorand-sdk)	111	117	Java	Cucumber	algorand/java-algorand-sdk (code and steps), algorand/algorand-sdk-testing (features)	MIT (both repositories)
coddy-agent (coddy-agent)	274	282	Go	godog	coddy-project/coddy-agent	MIT
CoreShop (coreshop)	334	422	PHP	Behat	coreshop/CoreShop	CoreShop Commercial License (CCL), with Pimcore under the Pimcore Open Core License (POCL)
e-petitions (e-petitions)	98	119	Ruby	Cucumber	alphagov/e-petitions	MIT
Eclipse Kapua (kapua)	330	605	Java	Cucumber	eclipse-kapua/kapua	EPL-2.0
PrestaShop (prestashop)	348	1,352	PHP	Behat	PrestaShop/PrestaShop	OSL-3.0 (core), AFL-3.0 (modules)
python-docx (python-docx)	237	239	Python	behave	python-openxml/python-docx	MIT
Sylius (sylius)	319	1,950	PHP	Behat	Sylius/Sylius	MIT
Total	2,051	5,086

Each project is pinned to one upstream commit (base_commit).

Dataset Structure

Data Fields

Default config (eval, full):

Field	Type	Description
instance_id	string	The stable key, <project>/<feature file path>:<Scenario line>.
project	string	One of algorand-sdk, coddy-agent, coreshop, e-petitions, kapua, prestashop, python-docx, sylius.
language	string	go, java, php, python or ruby.
framework	string	The BDD runner: godog, cucumber (Java and Ruby), behat or behave.
repo_url	string	The upstream repository of the project snapshot the test runs against.
base_commit	string	The 40-character commit of repo_url that pins the snapshot.
feature_repo_url	string	The repository holding the .feature file. It equals repo_url except for algorand-sdk, whose features live in algorand/algorand-sdk-testing.
feature_commit	string	The commit of feature_repo_url.
feature.file_path	string	The path of the .feature file in feature_repo_url.
feature.title	string	The text after Feature:.
feature.narrative	string or null	The Feature's free-text description as written. Null when the Feature has a title only.
feature.background	string or null	The Feature's Background: block, or null. Its Steps run before the Scenario's.
scenario.title	string	The text after Scenario: or Scenario Outline:.
scenario.line	int64	The 1-based line of the Scenario header in feature.file_path.
scenario.kind	string	scenario or outline. An Outline is one instance, and all its Examples rows run.
scenario.tags	list of string	The tags in effect for the Scenario (Feature and Scenario tags).
scenario.gherkin	string	The exact Scenario text from its header through its last Step or Examples row. With feature.* it forms the Requirement.
requirement_richness	string	The kind of Feature description: user_story, prose or title_only.
target_interfaces	list of {declaration, key}	The first-party public production methods that the ground-truth test calls directly from its When and Then Steps. declaration is the declaration text shown to agents. key identifies the method (2 or 3 strings, by language).
target_interfaces_empty_reason	string or null	Null when target_interfaces is non-empty. no_action_or_check_call when the action and check Steps call no production method directly. not_identified: ... for 6 rows of algorand-sdk outside eval.
test_level	string	component, integration or system.
test_level_class	string	The Step Definition suite class the level was assigned from, for example KP-UNIT.
in_eval_split	bool	True for the instances of eval.
gt_test_files	list of string	The paths (relative to repo_url at base_commit) of the ground-truth Test Implementation files. Their contents are in the gt_tests config.

gt_tests config:

Field	Type	Description
project	string	As above.
instance_id	string	Joins to the default config.
files	list of {path, content}	One entry per gt_test_files path, in the same order. content is the file's full UTF-8 text at base_commit.
di_fragment	string or null	CoreShop and Sylius only. The JSON text of the per-instance Symfony service definitions (service id, class, arguments, tags) that register the instance's Behat contexts. Null for the other projects.

Data Splits

Config	Split	Rows	Contents
default	eval	2,051	The benchmark instances. The metrics count 2,037 of them (see below).
default	full	5,086	Every execution-validated instance, eval included.
gt_tests	full	5,086	The ground-truth Test Implementation of every full instance. It is the reference answer, not model input.

The eval rows are the full rows with in_eval_split: true, in the same order. For every project except python-docx, full is analysis data that req2test does not evaluate.

The metrics leave out 14 CoreShop and Sylius eval instances whose target has no code to hide, or whose ground-truth test passes even with the target hidden (population).

Keep the reference data away from the agent under test. The gt_tests config, the base.tar.zst view archives and the kits hold the answer or the unhidden code. Give an agent only the materialized view of its instance.

Release Files

releases/<tag>/ holds the files that req2test fetch downloads and verifies (7.2 GB): the dataset as JSONL, and per project an evaluation kit, the source views and a runtime pack (docs/kits.md). The kits take 3.8 GB, including the MySQL image that the CoreShop and Sylius kits share. The runtime packs take 3.3 GB, the source views 59 MB and the dataset 1.5 MB. The dataset viewer and load_dataset do not read this folder.

How to Evaluate

The req2test command-line tool, its requirements and its installation are on GitHub. req2test fetch downloads and verifies the release files of this repository.

The GitHub usage guide shows how to prepare the tasks, run an agent on them and evaluate its predictions, with an example that runs Codex CLI. docs/agents.md describes how to run your own agent. Evaluation options, runs over several hosts and the cost per project are in docs/evaluation.md.

Licensing Information

The Req2Test-Bench code (GitHub) and the benchmark metadata authored for it (instance identifiers, target interfaces, labels, splits, mutant selections and kill sets) are released under the MIT licence.

Task content is derived from the upstream projects listed under Languages and stays under each project's own licence. This covers the requirements (Gherkin text), source snapshots and views, ground-truth tests, and mutants. The Docker images in the kits and the runtime packs contain third-party software (operating-system packages, language runtimes, build tools, databases such as MySQL Community Server and PostgreSQL, and each project's dependencies) under their own licences.

CoreShop content, including its kit, views and runtime pack, is under the CoreShop Commercial License and requires Pimcore under the Pimcore Open Core License. The terms are summarised in NOTICE.md.

Citation Information

@misc{req2testbench2026,
  title         = {Benchmarking {LLM} Agents for Behaviorally Adequate Test Generation from Software Requirements},
  author        = {Qi, Binhang and Liu, Yiming and Lin, Yun and Zhou, Yangjie and Sun, Hailong and Dong, Jin Song},
  year          = {2026},
  eprint        = {2610.36584},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE}
}