---
language:
  - en
pretty_name: BugWitGym
task_categories:
  - text-generation
tags:
  - code
  - java
  - software-engineering
  - coding-agents
  - bug-validation
  - executable-witness
  - test-generation
  - bug-injection
  - benchmark
  - llm-evaluation
size_categories:
  - 1K<n<10K
---

# BugWitGym: Evaluating Coding Agents for Executable Bug Validation

Anonymous dataset release accompanying the manuscript under double-blind review.

## Overview

BugWitGym evaluates coding agents on executable bug validation. Given a buggy repository, a bug-type description, and a reported bug location, an agent must construct an executable witness: a concrete input together with a testing harness that exposes the reported bug through observable faulty behavior.

The task is to validate an already-reported bug through execution, not to discover an unspecified bug or repair production code.

This repository contains the full benchmark release, a representative inspection sample, case-level metadata, Croissant metadata, and integrity checksums.

## Benchmark Summary

| Property | Description |
| --- | --- |
| Full benchmark | 1,300 execution-validated cases |
| Inspection sample | 51 representative cases |
| Source projects | 6 real-world Java projects |
| Build ecosystem | Maven |
| Bug types | 10 |
| Bug categories | API Contract, Value Flow, and Logic |
| Execution-context buckets | Short: 1–40; mid-short: 41–70; mid-long: 71–120; long: 121–150 distinct production-side functions |
| Transformation depths | 1, 3, 5, and 10 applied transformations |
| Evaluation settings | NoEC and WithEC |

The inspection sample is provided for reviewing the release structure and representative cases. It is not an additional set of cases beyond the full benchmark.

## Benchmark Construction

Cases are automatically constructed by injecting bugs into production-code paths exercised by existing tests. Each accepted case must contain a meaningful production-code change consistent with the selected bug type. The associated construction-time witness must pass on the clean baseline and expose the injected bug in the modified repository.

Bug-preserving transformations subsequently vary the surrounding call-stack, data-flow, and control-flow structure. A transformation is retained only when the project remains compilable and the same construction-time witness continues to expose the injected bug.

The resulting cases contain newly injected bugs embedded in real project contexts rather than a collection of previously disclosed historical bugs.

## Task and Evaluation Protocol

BugWitGym supports two evaluation settings:

| Setting | Agent-visible task information |
| --- | --- |
| NoEC | Buggy repository, bug-type description, and reported bug location, without an execution-context summary |
| WithEC | The same information, supplemented with the execution-context summary collected during construction |

Under the evaluation protocol, the agent creates or modifies Java test files under `src/test/java` to construct its own executable witness. Production code, build configuration, dependencies, and unrelated files must remain unchanged.

A successful attempt requires a materialized Java test that compiles and exposes the intended injected bug under the recorded Maven replay command. Compilation errors, missing dependencies, infrastructure failures, and unrelated test failures do not count as successful bug validation.

Release inspection and agent evaluation must be kept separate. Before evaluation, the construction-time witness must be removed or rewritten in the agent workspace. Construction-time witness implementations, hidden validation oracles, and construction-time solution artifacts must not be exposed to the evaluated agent.

## Repository Contents

| File | Contents |
| --- | --- |
| `pocgym_benchmark_release_v1.tar.gz` | Full 1,300-case benchmark release |
| `pocgym_benchmark_sample_v1.tar.gz` | Representative 51-case inspection sample |
| `pocgym_croissant.json` | Machine-readable Croissant dataset metadata |
| `CHECKSUMS.sha256` | SHA-256 checksums for uploaded artifacts |
| `MANIFEST.jsonl` | One machine-readable row per full-release case |
| `DATASET_SCHEMA.md` | Field definitions and metadata documentation |
| `LICENSES.md` | Dataset and upstream-source licensing notes |
| `SAMPLE_README.md` | Inspection-sample construction notes |
| `SAMPLE_SELECTION.json` | Selected sample case IDs and coverage reasons |

The existing `pocgym_` artifact filenames are retained for compatibility with the release layout. The benchmark described by this dataset card is BugWitGym.

## Per-Case Contents

| Path | Contents |
| --- | --- |
| `case.json` | Benchmark case metadata |
| `bug.patch` | Ground-truth patch documenting the injected bug |
| `buggy_repo/` | Cleaned full buggy repository workspace |

Case identifiers are globally sequential, beginning with `bj_case_000001`. This naming scheme supports appending additional cases without changing existing case identifiers or the directory layout.

Refer to `DATASET_SCHEMA.md` for the exact metadata fields and their interpretation.

## Inspecting the Release

Start with `SAMPLE_README.md`, `SAMPLE_SELECTION.json`, and the 51-case inspection archive to review representative cases. Use `MANIFEST.jsonl` and `DATASET_SCHEMA.md` to inspect the full-release index and metadata structure.

Verify downloaded artifacts against `CHECKSUMS.sha256` before extracting the archives.

The archives contain repository-level benchmark artifacts, not only prompt–completion records. Reproducing execution-based results requires a compatible Java runtime, Maven or Maven-wrapper support, and the case-specific replay environment described in the manuscript and case metadata.

## Scope and Limitations

The current release covers ten observable bug types in Maven-based Java projects with replayable tests. Its coverage should not be interpreted as representative of all programming languages, software ecosystems, or bug classes.

The benchmark is intended for evaluating executable witness construction. The representative sample is intended for inspection rather than as a substitute for the full benchmark or the evaluation subsets described in the manuscript.

Injected buggy repositories are benchmark artifacts and are not intended for production deployment.

## Licensing

See `LICENSES.md` for dataset and upstream-source licensing information. Consult the applicable license and notice files included with the bundled repositories before reuse.

## Anonymous Review

Author names, affiliations, and identifying project links are omitted from this dataset card for double-blind review.