---
license: gpl-3.0
---
# Juliet_LLVM Dataset

[![Hugging Face](https://img.shields.io/badge/🤗%20View%20on-HuggingFace-blue)](https://huggingface.co/datasets/compAgent/Juliet_LLVM)

## Dataset Summary

**Juliet_LLVM** is a dataset of **compiled C functions from the Juliet Test Suite** (as organized in the [GitHub repository](https://github.com/arichardson/juliet-test-suite-c)
), translated into **LLVM Intermediate Representation (IR)** after pre-process phase. It is designed for training and evaluating machine learning models on the task of **binary vulnerability detection**. Each function is labeled as either vulnerable or non-vulnerable and is presented in an architecture-agnostic, semantically rich format.

This LLVM version allows models to be trained and tested on realistic, compiler-transformed code — reflecting how such models would be deployed in real-world binary analysis scenarios.

This dataset supports experiments described in our paper and follows the same compilation, splitting, and evaluation procedures.

## Key Features

- ✅ Based on the **Juliet Test Suite**, a standard benchmark for vulnerability detection  
- ✅ **LLVM IR representation** of each function (field: `llvm_ir_function`)  
- ✅ Predefined **train**, **validation**, and **test** splits  
- ✅ **Binary vulnerability labels** (`label`) for classification  
- ✅ Includes metadata: original file and function name  
- ✅ Efficiently stored as **Parquet** files for fast loading and processing  

## Dataset Structure

Each record contains:

- `dataset`: The origin source, always `"Juliet"` for this dataset  
- `file`: Source file from the Juliet suite  
- `fun_name`: The name of the function  
- `llvm_ir_function`: The LLVM IR pre-processed code for the function  
- `label`: `"1"` for vulnerable, `"0"` for non-vulnerable  
- `split`: One of `"train"`, `"validation"`, or `"test"`  

## Split Information

This dataset is split into three subsets:

- `train`: Used for training the models  
- `validation`: Used for hyperparameter tuning and model selection  
- `test`: Held out for final evaluation and benchmarking  

> ✅ These splits match exactly the partitioning used in our **paper experiments**, allowing reproducibility and direct comparison with our reported results. Each split is disjoint and ensures no function-level overlap between sets.

## Format

This dataset is stored in [Apache Parquet](https://parquet.apache.org/) format under the `default` configuration. It adheres to the [Croissant schema](https://mlcommons.org/croissant/) and includes metadata for fields and splits.

## Usage

You can load the dataset using the Hugging Face 🤗 `datasets` library:

```python
from datasets import load_dataset

# Load train split
train_ds = load_dataset("compAgent/Juliet_LLVM", split="train")
print(train_ds[0])
```

```
{
  "dataset": "Juliet",
  "file": "CWE121/s01.c",
  "fun_name": "CWE121_bad",
  "llvm_ir_function": "define dso_local void @CWE121_bad() { ... }",
  "label": "1",
  "split": "train"
}
```

## License
This dataset is released under the GPL-3.0.

## Related Work
[Juliet Test Suite](https://samate.nist.gov/SARD/test-suites) — the original source of these functions.

[Juliet Test Suite GitHub repository](https://github.com/arichardson/juliet-test-suite-c) - the GitHub repository we took the Juliet Test Suite dataset from.


## Citation
```
@misc{juliet_llvm,
  author       = {Compote},
  title        = {Juliet_LLVM: A Dataset of Vulnerable and Non-Vulnerable Functions from the Juliet Suite in LLVM IR},
  howpublished = {\url{https://huggingface.co/datasets/compAgent/Juliet_LLVM}},
  year         = {2025}
}
```