26PaperSubmission/repo-vul

RepoVul Dataset

Repo-level Vulnerability Detection Dataset for Fine-tuning Lightweight LLMs

Dataset Description

RepoVul is a comprehensive dataset for repository-level vulnerability detection. This subset contains training, validation, and test data for fine-tuning lightweight Large Language Models (LLMs) on vulnerability detection tasks.

Key Features

• Task: Binary classification (vulnerable vs non-vulnerable code)
• Languages: Java and Python
• Format: Instruction-tuning prompt format
• Labels: 0 (non-vulnerable) or 1 (vulnerable)
• Context: Includes repository-level context with function call graphs

Data Format

Each example contains:

{
  "target": 1,  // 1 = vulnerable, 0 = non-vulnerable
  "func": "If the following code snippet has any vulnerabilities...",  // Prompt with code and context
  "idx": "0+train++CVE-2014-9635+vul.pkl",  // CVE mapping identifier
  "func_name_": "function_name"  // Function name
}

Field Descriptions

• target: Vulnerability label (1 = vulnerable, 0 = non-vulnerable)
• func: The complete prompt including instruction, code snippet, and repository context
• idx: Identifier used to map the function back to its CVE and repository
  • Format: {index}+{split}+{project_info}+{CVE_ID}+{label}.pkl
  • Purpose: This field identifies which CVE and whole repository this function belongs to
  • Important: Since multiple CVEs may share the same repository, we aggregate unique functions as test samples to avoid duplicate predictions. The idx field allows mapping predictions back to the original CVE/repository level results
• func_name_: The name of the function being analyzed

Prompt Structure

The func field includes:

1. Instruction: "If the following code snippet has any vulnerabilities..."
2. Code Snippet: The target function to analyze
3. Context Information:
   • File location
   • Functions that call this function (callers)
   • Functions called by this function (callees)

Dataset Splits

Java Data

Split	File	Size	Examples
Train	train.jsonl	53 MB	~200K
Validation	valid.jsonl	6.6 MB	~25K
Test	test.jsonl	220 MB	~800K

Python Data

Split	File	Size	Examples
Train	train_python.jsonl	22 MB	~80K
Validation	valid_python.jsonl	2.4 MB	~9K
Test	test_python.jsonl	94 MB	~350K

Example Usage

Java Dataset

import json

# Load Java training data
with open('train.jsonl', 'r') as f:
    for line in f:
        example = json.loads(line)
        print(f"Label: {example['target']}")
        print(f"Function: {example['func_name_']}")
        print(f"Prompt: {example['func'][:200]}...")
        break

Python Dataset

import json

# Load Python training data
with open('train_python.jsonl', 'r') as f:
    for line in f:
        example = json.loads(line)
        print(f"Label: {example['target']}")
        print(f"Function: {example['func_name_']}")
        print(f"Prompt: {example['func'][:200]}...")
        break

Mapping Predictions to CVE/Repository Level

Since multiple CVEs may share the same repository, predictions should be aggregated using the idx field:

import json
from collections import defaultdict

# Load test data and make predictions
predictions = {}  # function_name -> prediction
cve_mapping = defaultdict(list)  # cve_id -> list of function predictions

with open('test.jsonl', 'r') as f:
    for line in f:
        example = json.loads(line)

        # Your model makes prediction here
        pred = model.predict(example['func'])
        predictions[example['func_name_']] = pred

        # Parse idx to extract CVE ID
        # Format: "0+test+CVE-2011-4367+vul.pkl"
        parts = example['idx'].split('+')
        if len(parts) >= 4:
            cve_id = parts[2]  # Extract CVE ID
            cve_mapping[cve_id].append(pred)

# Aggregate predictions at CVE level
cve_results = {}
for cve_id, preds in cve_mapping.items():
    # Aggregate by majority vote or averaging
    cve_results[cve_id] = sum(preds) / len(preds)

print(f"CVE-level results: {cve_results}")

Use Cases

• Fine-tuning: Train lightweight LLMs (CodeBERT, GraphCodeBERT, etc.)
• Vulnerability Detection: Binary classification of code vulnerabilities
• Code Analysis: Repository-level vulnerability prediction
• Benchmarking: Evaluate model performance on vulnerability detection

Supported Vulnerability Types

The dataset covers various CWE categories including:

• CWE-79: Cross-site Scripting (XSS)
• CWE-89: SQL Injection
• CWE-119: Buffer Errors
• CWE-200: Information Exposure
• CWE-352: Cross-Site Request Forgery (CSRF)
• CWE-770: Allocation of Resources Without Limits
• CWE-20: Improper Input Validation
• And many more...

Sample Data

Java Example (Vulnerable)

private void updateSecureSessionFlag() {
    try {
        ServletContext context = Jenkins.getInstance().servletContext;
        Method m;
        try {
            m = context.getClass().getMethod("getSessionCookieConfig");
        } catch (NoSuchMethodException x) {
            LOGGER.log(Level.FINE, "Failed to set secure cookie flag", x);
            return;
        }
        Object sessionCookieConfig = m.invoke(context);

        Class scc = Class.forName("javax.servlet.SessionCookieConfig");
        Method setHttpOnly = scc.getMethod("setHttpOnly",boolean.class);
        setHttpOnly.invoke(sessionCookieConfig,true);

        Method setSecure = scc.getMethod("setSecure",boolean.class);
        boolean v = fixNull(jenkinsUrl).startsWith("https");
        setSecure.invoke(sessionCookieConfig,v);
    } catch (Exception e) {
        LOGGER.log(Level.WARNING, "Failed to set secure cookie flag", e);
    }
}

CVE: CVE-2014-9635 Label: Vulnerable

Python Example (Vulnerable)

def handle_get(self) -> bool:
    if self.do_log:
        logmsg = "%-4s %s @%s" % (self.mode, self.req, self.uname)

        if "range" in self.headers:
            try:
                rval = self.headers["range"].split("=", 1)[1]
            except:
                rval = self.headers["range"]

            logmsg += " [\033[36m" + rval + "\033[0m]"

        self.log(logmsg)

    # "embedded" resources
    if self.vpath.startswith(".cpr"):
        if self.vpath.startswith(".cpr/ico/"):
            return self.tx_ico(self.vpath.split("/")[-1], exact=True)

        # ... more code ...

CVE: CVE-2023-37474 Label: Vulnerable

License

This dataset is released under the MIT License.