VeriScope

VeriScope is an open benchmark dataset for first-pass Verilog coding.

This Hugging Face repository hosts the released benchmark bundles and dataset metadata for VeriScope.

Dataset summary

• domain: Verilog / RTL generation
• unit of release: task bundle
• target setting: first-pass code generation from natural-language task briefs
• public paper size: 568 problems

Dataset structure

Each public example is a task bundle containing:

• problem.md
• config.yaml
• testbench.v

The default public release excludes reference.v. The reference RTL is treated as a benchmark-side validation anchor rather than as part of the default evaluation package.

Suggested tree:

bundles/
├── L1_basic/
├── L2_sequential/
├── L3_module/
├── L4_system/
└── L5_ultimate/

Supported uses

• evaluation of Verilog code generation models
• evaluation of first-pass hardware coding agents
• analysis of functional correctness vs. structural quality

Out-of-scope uses

• production sign-off
• security certification
• claims of complete training-data decontamination

Dataset creation

The benchmark is built from a curated Verilog seed corpus and standardized into reusable problem bundles with benchmark-side validation harnesses. Public task briefs are rewritten and packaged rather than copied verbatim from source material.

Evaluation

The companion GitHub repository contains the public evaluation harness, scoring logic, and reproduction instructions.

• code repository: https://github.com/zwpride/VeriScope
• dataset repository: https://huggingface.co/datasets/zwpride/VeriScope

Licensing

The released benchmark bundles are distributed under CC BY 4.0. This license applies to the released task packages in this repository. Third-party source material used during benchmark construction is not redistributed verbatim here and remains under its original terms.

Additional metadata

• Croissant metadata: croissant.json
• responsible release notes: RAI.md
• dataset layout notes: DATASET_STRUCTURE.md
• dataset license note: LICENSE.md