---
annotations_creators:
  - found
language_creators:
  - found
language:
  - en
license: mit
multilinguality:
  - monolingual
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - text-generation
  - question-answering
task_ids:
  - language-modeling
  - language-modeling
pretty_name: Bug Bounty Hunter
tags:
  - cybersecurity
  - vulnerability
  - bug-bounty
  - security
  - instruction-tuning
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-00000-of-00001.parquet
      - split: validation
        path: data/validation-00000-of-00001.parquet
      - split: test
        path: data/test-00000-of-00001.parquet
---

# Bug Bounty Hunter v1 - Fine-Tuning Dataset

The **largest public dataset for training AI models on bug bounty hunting expertise**.

## Overview

52,743 instruction-response pairs covering 40+ vulnerability types across 6 categories, extracted from real-world bug bounty writeups and security research.

### Dataset Summary

This dataset is designed to fine-tune language models to become expert bug bounty hunters. Each example contains a technical instruction and a detailed response drawn from thousands of real vulnerability writeups, security blogs, CVE analyses, and methodology guides.

### Use Cases

- **Fine-tune chat models** for bug bounty expertise
- **Train vulnerability classifiers** 
- **Teach LLMs security testing methodology**
- **Build automated security tools**

### Dataset Structure

| Split | Examples | Description |
|-------|----------|-------------|
| `train` | 42,194 | Training set (80%) |
| `validation` | 5,274 | Validation set (10%) |
| `test` | 5,275 | Test set (10%) |

### Data Format (ChatML)

```json
{
  "messages": [
    {"role": "system", "content": "You are an expert bug bounty hunter..."},
    {"role": "user", "content": "Explain XSS in detail."},
    {"role": "assistant", "content": "Cross-Site Scripting (XSS) is a vulnerability..."}
  ]
}
```

### Instruction Types

| Type | Count | Description |
|------|-------|-------------|
| explain | 10,165 | Vulnerability explanations |
| find | 10,157 | Detection techniques |
| exploit | 10,149 | Exploitation methods |
| prevent | 10,138 | Prevention & mitigation |
| impact | 10,134 | Business impact analysis |
| writeup_analysis | 2,000 | Bug report analysis |

### Vulnerability Types (40+)

**Web Exploitation:** XSS, SQL Injection, SSRF, CSRF, XXE, IDOR, Path Traversal, Command Injection, Open Redirect, File Upload, Subdomain Takeover, Clickjacking

**API & Modern:** GraphQL, API Misconfigurations, Mass Assignment, NoSQL Injection, JWT Attacks, OAuth Flaws, Cache Attacks, Host Header

**Advanced:** RCE, Deserialization, Template Injection, Request Smuggling, Race Conditions, Business Logic

**Reconnaissance:** Subdomain Discovery, Content Discovery, JS Source Analysis, Port Scanning, Technology Fingerprinting, Cloud Recon

**Methodology:** Burp Suite, Automation, Mobile Testing, Cloud Testing, IoT Firmware

### Load with HuggingFace

```python
from datasets import load_dataset

dataset = load_dataset("your-username/bugbounty-hunter-v1")
print(dataset)
# DatasetDict({
#     train: Dataset({ features: ['system', 'user', 'assistant', 'messages_json'], num_rows: 42194 })
#     validation: Dataset({ features: [...], num_rows: 5274 })
#     test: Dataset({ features: [...], num_rows: 5275 })
# })

# Format for chat models
def format_chat(example):
    return {"text": f"System: {example['system']}\nUser: {example['user']}\nAssistant: {example['assistant']}"}
```

### Load with Transformers

```python
from transformers import AutoTokenizer
from datasets import load_dataset

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
dataset = load_dataset("your-username/bugbounty-hunter-v1")

def format_for_training(examples):
    texts = []
    for sys, user, asst in zip(examples["system"], examples["user"], examples["assistant"]):
        messages = [
            {"role": "system", "content": sys},
            {"role": "user", "content": user},
            {"role": "assistant", "content": asst},
        ]
        texts.append(tokenizer.apply_chat_template(messages, tokenize=False))
    return tokenizer(texts, truncation=True, max_length=2048)

dataset = dataset.map(format_for_training, batched=True)
```

### Source Data

Built from a massive collection of 6.32 GB / 17,014 files:
- 200,000 CVEs from NVD
- 4,200+ full bug bounty writeups (HTML)
- 6,421 PentesterLand indexed writeups
- 3,300+ HackerOne reports
- 35+ security research blogs
- 164+ wordlists (323+ MB)
- OWASP guides, exploit databases, CTF writeups

### Model Training

For training, see our training scripts:
- LoRA with Unsloth: `Training/train_unsloth.py`
- QLoRA with HF: `Training/qlora_config.json`
- Full fine-tune: `Training/full_ft_config.json`

### Benchmark

Evaluate your model with `AI-Datasets/bug_bounty_benchmark.json` (102 curated questions across 20+ categories).

### Citation

```bibtex
@dataset{bugbounty-hunter-v1,
  title        = {Bug Bounty Hunter v1 - Fine-Tuning Dataset},
  author       = {Bug Bounty Community},
  year         = {2026},
  description  = {52,743 instruction-response pairs for bug bounty hunting expertise},
  url          = {https://huggingface.co/datasets/your-username/bugbounty-hunter-v1},
  license      = {MIT}
}
```