Binary-30K: Cross-Platform Binary Dataset with Stratified Splits

Paper | Code

🔗 Original Dataset (no splits): mjbommar/binary-30k-tokenized

This is the stratified train/validation/test split version of the Binary-30K dataset, containing 29,793 unique cross-platform binaries with pre-computed tokenization. This version provides standardized splits for reproducible machine learning research.

🎯 Key Features

• ✅ Stratified 70/15/15 splits maintaining class balance across all sets
• ✅ 4-dimensional stratification across malware/platform/format/architecture
• ✅ 26.9% malware balance preserved in all splits (±0.1%)
• ✅ Deterministic splits (seed=42) for reproducible research
• ✅ Ready for ML - no manual splitting required
• ✅ Pre-computed BPE tokenization for transformer models

📊 Dataset Splits

Split	Samples	Malware	Benign	Malware %
Train	20,849	5,613	15,236	26.92%
Validation	4,463	1,200	3,263	26.89%
Test	4,481	1,210	3,271	27.00%
Total	29,793	8,023	21,770	26.93%

Stratification Strategy

Splits maintain proportional representation across:

• ✅ Malware vs. Benign (26.9% malware in each split)
• ✅ Platform (Windows, Linux, macOS, Android, Other)
• ✅ File Format (PE, ELF, Mach-O, APK)
• ✅ Architecture Groups (common: x86/ARM vs. exotic: MIPS/RISC-V/PowerPC)

19 unique strata identified with proportional representation maintained across all splits.

🚀 Quick Start

from datasets import load_dataset

# Load dataset with splits
dataset = load_dataset("mjbommar/binary-30k")

train_ds = dataset["train"]        # 20,849 samples
val_ds = dataset["validation"]      # 4,463 samples
test_ds = dataset["test"]           # 4,481 samples

# Access pre-computed tokens
sample = train_ds[0]
print(f"Platform: {sample['platform']}")
print(f"Malware: {sample['is_malware']}")
print(f"Tokens: {len(sample['tokens'])} tokens")

Example: Malware Classification

from datasets import load_dataset
from transformers import Trainer, TrainingArguments

# Load data
dataset = load_dataset("mjbommar/binary-30k")

# Tokens are pre-computed - just truncate
def prepare_example(example):
    return {
        "input_ids": example["tokens"][:512],
        "labels": int(example["is_malware"])
    }

# Train on standard splits
train_ds = dataset["train"].map(prepare_example)
val_ds = dataset["validation"].map(prepare_example)

# Train your model...

Example: Cross-Platform Transfer Learning

# Train on Windows, test on Linux
train_windows = dataset["train"].filter(lambda x: x["platform"] == "windows")
test_linux = dataset["test"].filter(lambda x: x["platform"] == "linux")

print(f"Windows training samples: {len(train_windows)}")
print(f"Linux test samples: {len(test_linux)}")
# Evaluate cross-platform generalization...

📦 Dataset Composition

Platform Distribution:

• Windows: 57.3% (17,239 samples) - PE32/PE32+ executables and DLLs
• Linux: 28.4% (8,452 samples) - ELF32/ELF64 from 9 distributions
• macOS: 1.9% (568 samples) - x86-64, ARM64, Universal binaries
• Android: 0.6% (164 samples) - APKs with native ARM libraries
• Other: 11.8% (3,370 samples) - Diverse formats and installers

Architecture Diversity:

• Common: x86-64 (56.4%), x86 (11.1%), ARM64 (5.9%), ARM (9.4%)
• Exotic: MIPS (2.3%), PowerPC (1.3%), RISC-V (0.1%), m68k, SuperH, ARCompact, SPARC, S/390

Malware Sources:

• SOREL-20M: 365 Windows PE malware samples (2017-2019)
• Malware Bazaar: 7,658 cross-platform malware samples (2020-2024)
  • Platform-first stratified sampling
  • ALL available macOS malware (560 samples)
  • ALL available Android malware (164 samples)
  • Balanced Windows/Linux with size stratification

📋 Data Structure

Each record contains 31 fields organized into seven categories:

Identification (6 fields):

• file_id, file_path, file_name, sha256, md5, file_size

Platform/Source (5 fields):

• platform, os_family, os_version, distribution, is_malware

File Characteristics (6 fields):

• file_format, architecture, binary_type, is_stripped, is_packed, is_signed

Structural Analysis (4 fields + sections):

• num_sections, code_size, data_size, sections[]

Dependencies (4 fields + imports/exports):

• num_imports, num_exports, imports[], exports[]

Complexity (1 field):

• entropy (Shannon entropy 0-8 scale)

Pre-computed Tokenization (4 fields):

• tokens[], token_count, compression_ratio, unique_tokens

Parser Diagnostics (2 fields):

• parse_status, parse_warnings[]

Pre-computed Tokenization

All binaries are tokenized using BPE tokenization (mjbommar/binary-tokenizer-001-64k):

• Average tokens per binary: ~15,000
• Compression ratio: ~4.2 bytes/token
• Vocabulary: 64K tokens
• Ready for transformers: BERT, GPT, T5, etc.

🎓 Supported Research Tasks

1. Malware Detection: Binary classification with balanced classes (26.9% malware)
2. Cross-Platform Analysis: Transfer learning across Windows/Linux/macOS/Android
3. Architecture-Invariant Detection: Generalization to exotic architectures (IoT/embedded)
4. Mobile Malware Research: Dedicated Android and macOS malware samples
5. Binary Similarity: Embedding learning for similar binary detection
6. Format-Agnostic Analysis: Multi-format models (PE/ELF/Mach-O/APK)

📊 Comparison with Other Datasets

| Dataset | Size | Platforms | Architectures | Malware | Pre-tokenized | Splits | |---------|------|-----------|---------------|---------|---------------|--------|
| Binary-30K | 30K | Win+Linux+macOS+Android | 15+ (incl. exotic) | 26.9% | ✅ | ✅ | | SOREL-20M | 20M | Windows only | x86/x64 | 100% | ❌ | ❌ | | EMBER | 1.1M | Windows only | x86/x64 | 50% | ❌ (features) | ✅ | | Assemblage | 1.1M | Windows+Linux | x86/x64 | 0% (benign) | ❌ | ❌ |

🔍 Stratification Verification

Split Distribution Verification:

TRAIN (20,849 samples):

• Malware: 5,613 (26.92%)
• Top platforms: Windows (12,065), Linux (5,915), Other (1,200)
• Top formats: PE (12,018), ELF (5,915), Unknown (1,195)

VALIDATION (4,463 samples):

• Malware: 1,200 (26.89%)
• Top platforms: Windows (2,584), Linux (1,266), Other (256)
• Top formats: PE (2,574), ELF (1,266), Unknown (255)

TEST (4,481 samples):

• Malware: 1,210 (27.00%)
• Top platforms: Windows (2,590), Linux (1,271), Other (259)
• Top formats: PE (2,577), ELF (1,271), Unknown (258)

Statistical Tests: Chi-square tests confirm no significant deviation from proportional representation (p > 0.05 for all dimensions).

🔄 Reproducibility

Split Generation:

• Seed: 42 (for reproducibility)
• Method: Stratified sampling with composite keys
• Date: November 15, 2025
• Tool: binary-dataset-paper

All splits are deterministic and reproducible. Using the same seed will always produce identical splits.

📚 Data Sources

Linux Binaries: Alpine 3.18/3.19, Debian 11-12, Ubuntu 20.04/22.04/24.04, Fedora 39-40, CentOS Stream 9, Arch Linux, Kali Linux 2024.1, Parrot OS 6.0, BusyBox 1.37.0

Windows Binaries: Windows 8 Pro, Windows 10 21H2/22H2, Windows 11 23H2, Windows Update Catalog

Malware Samples:

• SOREL-20M dataset (Sophos-ReversingLabs, 2020)
• Malware Bazaar (abuse.ch, 2020-2024) with platform-first stratified sampling

⚠️ Important Considerations

Limitations:

• Static analysis only (no dynamic/runtime behavior)
• Some binaries cannot be parsed by LIEF
• Many binaries have stripped debug symbols
• Very large binaries produce extended token sequences
• iOS/iPadOS binaries not included
• Uneven representation of exotic architectures

Usage Notes:

• Malware samples require secure, isolated research environments
• Windows binaries subject to Microsoft licensing terms
• Fair use application depends on jurisdiction
• Splits are standardized but users may create custom splits for specific research needs

📄 License and Attribution

Dataset Compilation: CC-BY-4.0 license by Michael J. Bommarito II

Component Licenses:

• Linux binaries: Various open-source licenses (GPL, LGPL, MIT, BSD, Apache)
• Windows binaries: Subject to Microsoft software licenses
• SOREL-20M samples: Follow SOREL-20M License Agreement
• Malware Bazaar samples: Research use only, attribution required to abuse.ch

Malware samples are included for research purposes only. Users must comply with applicable laws and regulations when working with malware samples.

📖 Citation

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

@dataset{bommarito2025binary30k,
  title={Binary-30K: A Cross-Platform, Multi-Architecture Binary Dataset with Stratified Splits},
  author={Bommarito, Michael J., II},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/mjbommar/binary-30k}
}

🔗 Related Resources

• Original dataset (no splits): mjbommar/binary-30k-tokenized
• Tokenizer: mjbommar/binary-tokenizer-001-64k
• Paper: Binary-30K: A Heterogeneous Dataset for Deep Learning in Binary Analysis and Malware Detection
• Code & Documentation: github.com/mjbommar/binary-dataset-paper
• Technical Documentation: See DATASET_SPLITS.md for detailed stratification methodology

📞 Contact

Author: Michael J. Bommarito II Email: michael.bommarito@gmail.com

🔄 Updates

• 2025-11-15: Initial release with stratified train/val/test splits (70/15/15)

---

Last Updated: November 15, 2025