VeriScope / RAI.md
zwpride's picture
Upload RAI.md with huggingface_hub
e5c54d0 verified

Responsible AI and Release Notes

This file documents the responsible release notes for the public VeriScope dataset.

Resource type

  • benchmark dataset
  • evaluation suite companion data release

Intended use

  • evaluation of first-pass Verilog code generation
  • comparison of hardware coding models under a common benchmark
  • study of execution-only vs. quality-aware evaluation

Foreseeable misuse

  • generation of unsafe or malicious RTL
  • unauthorized recreation of proprietary-style hardware patterns
  • overclaiming model readiness for production chip design

Limitations

  • public third-party models are not fully decontaminated
  • top-end expert-scale coverage is still limited
  • benchmark scores do not replace synthesis, timing, formal verification, or human sign-off
  • the default public dataset excludes reference.v, so benchmark-side gold verification is not fully reproduced from the dataset release alone

Human oversight

The benchmark should be used as an evaluation aid, not as a final arbiter of production hardware quality.

Release safeguards

  • clear licensing and attribution
  • explicit statement of intended use
  • explicit statement of out-of-scope uses
  • public documentation of benchmark limitations
  • default exclusion of benchmark-side reference RTL from the public dataset