README.md

metadata

dataset_info:
  features:
    - name: description
      dtype: string
    - name: query
      dtype: string
    - name: source
      dtype: string
    - name: schema
      dtype: string
    - name: id
      dtype: int64
    - name: query_length
      dtype: int64
    - name: description_length
      dtype: int64
    - name: complexity_score
      dtype: int64
    - name: query_type
      dtype: string
    - name: entities
      sequence: string
  splits:
    - name: train
      num_bytes: 3580759
      num_examples: 301
  download_size: 33073
  dataset_size: 3580759
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

HoundBench Cypher Queries Dataset

Dataset Description

This dataset contains 180 curated Cypher queries specifically designed for BloodHound, the popular Active Directory reconnaissance tool. Each entry pairs a natural language description with its corresponding Cypher query, train and eval your agents for BloodHound query generation :D.

Dataset Summary

• Total Examples: 180 query-description pairs
• Language: English (descriptions), Cypher (queries)
• Domain: Cybersecurity, Active Directory analysis, Graph databases
• Use Cases: Query generation, cybersecurity education, BloodHound automation

Supported Tasks

• Text-to-Code Generation: Generate Cypher queries from natural language descriptions
• Query Understanding: Understand the intent behind cybersecurity queries
• Educational Resource: Learn BloodHound query patterns and techniques

Dataset Structure

Data Instances

Each example contains:

{
  "description": "Find all users with an SPN (Kerberoastable users)",
  "query": "MATCH (n:User) WHERE n.hasspn=true RETURN n",
  "source": "https://hausec.com/2019/09/09/bloodhound-cypher-cheatsheet/"
}

Data Fields

• description (string): Natural language description of what the query accomplishes
• query (string): The corresponding Cypher query for BloodHound/Neo4j
• source (string): Attribution to the original source (URL, author, or publication)

Data Splits

The dataset is provided as a single collection. Users can create custom splits using the provided utilities:

from datasets import load_dataset
from utils.dataset_utils import split_dataset

dataset = load_dataset("joshtmerrill/HoundBench")
train_set, test_set = split_dataset(dataset, train_ratio=0.8)

Additional Information

Dataset Curators

This dataset was curated as part of the HoundBench project, a comprehensive toolkit for evaluating and validating Cypher queries against BloodHound instances.

Queries were curated from open and closed sources.

Licensing Information

This dataset is released under the MIT License. While the dataset itself is freely available, users should respect the original sources and their respective licenses.

Citation Information

If you use this dataset in your research, please cite:

@dataset{houndbench,
  title={HoundBench: Benchmarking offensive agents},
  author={Josh Merrill},
  year={2025},
  url={https://huggingface.co/datasets/joshtmerrill/HoundBench},
}

Contributions

We welcome contributions to improve and expand this dataset. Please see our contribution guidelines for more information.

Usage Examples

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("joshtmerrill/bloodhound-cypher-queries")

# Load with custom split
train_dataset = load_dataset("joshtmerrill/bloodhound-cypher-queries", split="train[:80%]")
test_dataset = load_dataset("joshtmerrill/bloodhound-cypher-queries", split="train[80%:]")

Basic Usage

# Iterate through examples
for example in dataset:
    print(f"Description: {example['description']}")
    print(f"Query: {example['query']}")
    print(f"Source: {example['source']}")
    print("---")

Integration with HoundBench

from utils.dataset_utils import load_queries_dataset, split_dataset

# Load using HoundBench utilities
dataset = load_queries_dataset("joshtmerrill/bloodhound-cypher-queries")

# Create train/test split
train_set, test_set = split_dataset(dataset, train_ratio=0.8, random_seed=42)

# Filter by source
hausec_queries = filter_dataset_by_source(dataset, ["hausec.com"])

Query Generation Example

from transformers import pipeline

# Load a text generation model
generator = pipeline("text-generation", model="your-model")

# Generate query from description
description = "Find all Domain Admins with active sessions"
prompt = f"Description: {description}\nQuery:"
result = generator(prompt, max_length=100)
print(result[0]['generated_text'])