dilekmandal/owasp-benchmark-java-security

OWASP Benchmark Java – LLM Fine-Tuning Dataset

A chat-formatted dataset derived from the OWASP Benchmark for Java and prepared for large language model fine-tuning for Java vulnerability detection.

Dataset Summary

This dataset contains 21,041 Java security analysis examples derived from OWASP Benchmark test cases.

Each example contains:

• A system instruction describing the security analysis task
• A Java source-code sample
• An assistant response containing:
  • vulnerable: vulnerability presence (true / false)
  • cwe: CWE identifier
  • category: vulnerability category

The data is formatted using the messages structure commonly used for conversational and instruction fine-tuning.

Dataset Structure

Split	Examples
Train	16,849
Validation	2,096
Test	2,096
Total	21,041

Example

{
  "messages": [
    {
      "role": "system",
      "content": "You are a Java application security analyzer..."
    },
    {
      "role": "user",
      "content": "Analyze the following Java source code..."
    },
    {
      "role": "assistant",
      "content": "{\"vulnerable\": true, \"cwe\": \"CWE-501\", \"category\": \"trustbound\"}"
    }
  ]
}

Labels

The dataset contains the following CWE categories:

• CWE-22 – Path Traversal
• CWE-78 – OS Command Injection
• CWE-79 – Cross-Site Scripting
• CWE-89 – SQL Injection
• CWE-90 – LDAP Injection
• CWE-327 – Broken/Risky Cryptographic Algorithm
• CWE-328 – Weak Hashing
• CWE-330 – Weak Randomness
• CWE-501 – Trust Boundary Violation
• CWE-614 – Sensitive Cookie Without 'Secure' Flag
• CWE-643 – XPath Injection

The primary binary classification target is:

vulnerable = true / false

Intended Use

The dataset is intended for:

• Java source-code vulnerability detection
• Security-focused LLM fine-tuning
• Research on code security classification
• Evaluation of vulnerability-aware language models
• CWE classification experiments

Data Source

The underlying source is the OWASP Benchmark for Java, an open-source benchmark designed to evaluate the accuracy, coverage, and speed of software vulnerability detection tools.

The OWASP Benchmark project states that its test cases contain intentionally vulnerable and non-vulnerable examples mapped to specific CWEs.

Source project:

https://github.com/OWASP-Benchmark/BenchmarkJava

Project information:

https://owasp.org/www-project-benchmark/

Dataset Preparation

The original OWASP Benchmark Java examples were transformed into a conversational JSONL format suitable for LLM fine-tuning.

The transformation:

1. Extracted the Java source code from each benchmark test.
2. Associated each example with its corresponding CWE.
3. Preserved the vulnerable/non-vulnerable label.
4. Preserved the vulnerability category.
5. Converted each example into a system / user / assistant message structure.
6. Created train, validation and test splits.

No model-generated labels were used.

Split

The dataset contains:

• train.jsonl
• validation.jsonl
• test.jsonl

The current split contains approximately 80% training data, 10% validation data and 10% test data.

Distribution

The CWE distribution is approximately consistent across the three splits.

The vulnerable/non-vulnerable distribution is also approximately consistent across the splits.

Important Note About Benchmark Similarity

OWASP Benchmark is a synthetic security benchmark containing many test cases based on common Java application structures and vulnerability patterns.

Consequently, structurally similar source-code examples may occur across different splits.

This dataset should therefore be interpreted as an OWASP Benchmark-based fine-tuning dataset rather than as a leakage-controlled real-world software corpus.

For rigorous generalization studies, evaluation on an independent real-world Java vulnerability dataset is recommended.

Limitations

This dataset is based on a synthetic benchmark rather than naturally occurring production vulnerabilities.

Performance on this dataset should not be interpreted as equivalent to performance on real-world enterprise Java applications.

Models trained on this dataset may learn benchmark-specific code patterns.

For this reason, results should preferably be reported together with evaluation on additional independent datasets.

License

The underlying OWASP Benchmark for Java is distributed under the GNU General Public License v2.0 (GPL-2.0).

Users should review the original project's license and terms before redistributing or using derivative datasets.

The original project and its license information are available at:

https://github.com/OWASP-Benchmark/BenchmarkJava

Attribution

This dataset is derived from the OWASP Benchmark Project.

OWASP Benchmark Project: https://owasp.org/www-project-benchmark/

OWASP Benchmark for Java: https://github.com/OWASP-Benchmark/BenchmarkJava

This repository does not claim ownership of the original OWASP Benchmark source material.

Citation

If you use this dataset, please cite the original OWASP Benchmark project in addition to citing this dataset repository.

Disclaimer

This dataset is intended for research and security engineering purposes.

The presence of a vulnerability label reflects the OWASP Benchmark's intended test classification and does not constitute a security assessment of production software.