README.md

---
license: gpl-3.0
---
# CompRealVul_LLVM Dataset

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

## Dataset Summary

**CompRealVul_LLVM** is the LLVM IR (Intermediate Representation) version of the [CompRealVul_C](https://huggingface.co/datasets/CCompote/CompRealVul_C) dataset. This version is designed specifically for **training and evaluating machine learning models** on the task of **binary vulnerability detection** in a setting that closely mimics how models are used in practice — operating on the compiled representation of code rather than raw source code.

Each function in this dataset was compiled from C code to LLVM IR, enabling robust training of models on semantically rich, architecture-independent binary representations.

This dataset supports research aligned with the methodology described in our paper, where the goal is to predict vulnerabilities directly from **compiled IR representations** rather than from source code.

## Key Features

- ✅ **LLVM IR representation** of each function (field: `llvm_ir_function`)
- ✅ Includes **train**, **validation**, and **test** splits (see below)
- ✅ Vulnerability labels (`label`) for supervised learning
- ✅ Metadata about original source (`dataset`, `file`, `fun_name`)

## Dataset Structure

Each record contains:

- `dataset`: Original dataset source of the function (e.g., Juliet, NVD)
- `file`: File path of the source from which the function was extracted
- `fun_name`: Name of the function in the source code
- `llvm_ir_function`: LLVM IR string representing the function
- `label`: Binary label indicating vulnerability (`1` for vulnerable, `0` for non-vulnerable)
- `split`: Dataset split (`train`, `validation`, `test`)

## Split Information

This dataset is split into **train**, **validation**, and **test** sets, following the exact partitioning strategy used in the experiments described in our paper. The split ensures a fair evaluation of generalization performance by separating functions into disjoint sets with no overlap. This allows researchers to directly reproduce our results or compare against them under consistent conditions.

- `train`: Used to fit model parameters
- `validation`: Used for model selection and hyperparameter tuning
- `test`: Used exclusively for final evaluation and benchmarking


## Usage

You can load and explore the dataset using the 🤗 `datasets` library:

```python
from datasets import load_dataset

# Load a specific split
train_ds = load_dataset("compAgent/CompRealVul_LLVM", split="train")
print(train_ds[0])

## Example
```json
{
  "dataset": "CompRealVul",
  "file": "app_122.c",
  "fun_name": "app",
  "llvm_ir_function": "define dso_local i32 @app() #0 { ... }",
  "label": "1",
  "split": "train"
}
```

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

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