Open-source Hardware Debug(OSHD) Dataset

Major contributors:

Mr. Weimin Fu (Kansas State University)
Dr. Kaichen Yang (Michigan Technological University)
Dr. Raj Gautam Dutta (Silicon Assurance)
Dr. Xiaolong Guo (Kansas State University)
Dr. Gang Qu (University of Maryland)

✉ Correspondence to: Weimin Fu (weiminf@ksu.edu) or Xiaolong Guo (guoxiaolong@ksu.edu)

Citation Information

Please cite the following paper when using the OSHD Dataset.

@article{fu2023llm4sechw,
  title={LLM4SecHW: Leavering Domain-Specific Large Language Model for Hardware Debugging},
  author={Fu, Weimin and Yang, Kaichen and Dutta, Raj Gautam and Guo, Xiaolong and Qu, Gang},
  journal={Asian Hardware Oriented Security and Trust (AsianHOST)},
  year={2023}
}

arxiv: 2401.16448

Update from our group for Hardware domain-specific LLM:

Blog: Large Language Model for Hardware Security

HomePage: Hardware Security Lab

Introduction

The Open-source Hardware Debug Dataset is a meticulously curated collection of information extracted and cleansed from the version control systems of various popular open-source hardware designs. This dataset is designed to aid in the development and fine-tuning of Language Models (LLMs) for hardware debugging tasks.

Dataset Utilization in Research

This dataset has been prominently utilized in the research paper:

Fu, W., Yang, K., Dutta, R. G., Guo, X., & Qu, G. (2023). LLM4SecHW: Leveraging Domain-Specific Large Language Model for Hardware Debugging. Asian Hardware Oriented Security and Trust (AsianHOST). You can view the paper from arxiv: arxiv.org/abs/2401.16448

Dataset Composition

Our dataset includes data from several renowned open-source hardware projects, such as:

Pulp: GitHub - Pulp Platform
darkriscv: GitHub - darklife
OpenPiton: GitHub - Princeton University
CVA6: GitHub - OpenHW Group
mor1kx: GitHub - OpenRISC
Ibex: GitHub - lowRISC
OpenTitan: GitHub - lowRISC

Building Process

The dataset building starts with issues labeled as 'Bug'. By analyzing the dialogue within these issues and identifying associated pull requests (PRs), we locate the PR corresponding to each issue. Each PR includes 'Files changed' information, enabling us to identify the modified hardware design source code, its path, the version before and after the modification, and the modification itself (patch).

Supported Tasks

This dataset is particularly tailored for fine-tuning Language Learning Models (LLMs) in hardware debugging tasks.

Dataset Structure

The dataset is available in two versions:

Version 1 includes modifications in hardware projects written in Verilog language.
Version 2 encompasses code modifications marked as BUG fixes, traced through the complete issue-PR-file change pathway.

Examples

Version 1 Example:

{
  "filename": "bhv/cv32e40s_rvfi.sv",
  "before": "// Copyright (c) 2020 OpenHW Group\n//\n//.....",
  "after": "// Copyright (c) 2020 OpenHW Group\n//\n//.....",
  "commit_message": "Merge pull request #468 from silabs-oysteink/s",
  "pull_number": "468",
  "pull_title": "Fix for issue where rvfi_intr could have the w...",
  "pull_body": "Resetting 'dummy_suppressed_intr' if a new int..."
}

Version 2 Example:

{
  "filename": "corev_apu/tb/ariane_...",
  "before": "/* Copyright 2018 ET...",
  "after": "/* Copyright 2018 ET...",
  "commit_message": "AXI master in Ariane...",
  "pull_number": 1182,
  "pull_title": "Add AXI parameter in...",
  "pull_body": "The axi stuctures we...",
  "repo_name": "openhwgroup/cva6",
  "commit_sha": "3919983fdf96b152a877...",
  "patch": "@@ -17,15 +17,12 @@\n...",
  "Issue title": "[BUG] rvfi_pkg dupli...",
  "Issue Body": "### Is there an exis...",
  "Issue labels": "Type:Bug",
  "Comments": "JeanRochCoulon: Good..."
}

Data Fields

The Open-source Hardware Debug Dataset consists of the following fields in its JSON structure, providing comprehensive information about each code modification:

Filename: The path and name of the file that was modified in the hardware project.
Before: The content of the file before the modification, typically showing the original code or configuration.
After: The content of the file after the modification, illustrating the changes made.
Commit Message: A brief description of the changes made, usually associated with the commit in version control systems.
Pull Number: The unique identifier of the pull request on the project’s repository that contains the modification.
Pull Title: The title of the pull request, often summarizing the nature or purpose of the change.
Pull Body: A detailed description provided in the pull request, explaining the reason for the change, its impact, and any additional context.
Repo Name: The name of the repository where the modification occurred, providing a reference to the source project.
Commit SHA: The SHA (Secure Hash Algorithm) identifier of the commit, offering a precise reference to the specific change in the project’s history.
Patch: A detailed view of the changes made in the file, represented in diff format, showing exactly what was added, removed, or altered.
Issue Title: The title of the issue related to the bug that the modification addresses.
Issue Body: The detailed description of the issue, providing insights into the problem that prompted the modification.
Issue Labels: Tags or labels assigned to the issue, such as "Type:Bug," helping to categorize and prioritize the issue.
Comments: Any comments made on the issue, offering additional perspectives or discussions related to the modification.

This comprehensive set of data fields ensures that users of the dataset have a thorough understanding of each modification, its context, and its implications, making it an invaluable resource for debugging and improving hardware designs.

Data Splits

Each configuration contains a single training split.

Acknowledgment

Portions of this work were supported by the National Science Foundation (CCF-2019310, First Award Program of ARISE in EPSCoR 2148878).

License

This dataset is made available under the Apache License, Version 2.0 (Apache-2.0). The Apache 2.0 License is a permissive open-source license that allows for broad use, modification, and distribution of the licensed software.

Key points of the Apache 2.0 License include:

Permissiveness: This license allows for the use, reproduction, modification, distribution, and sublicensing of the software, both in source and binary forms.
Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made.
State Changes: If you modify the software, you must clearly indicate your changes.
No Warranty: The software is provided "as is" without any warranties.

For the full terms and conditions, as well as further details, please refer to the Apache License, Version 2.0 official webpage.

Datasets:

KSU-HW-SEC
/

LLM4SecHW-OSHD

You need to agree to share your contact information to access this dataset