Arrakark/scalared_guidelines

---
license: gpl-3.0
language:
- en
tags:
- digital
- design
- verificaiton
- systemverilog
- verilog
pretty_name: scalared
---

# Scalared Guidelines
## Description
The following dataset is a collection of 7226 rules and guidelines related to digital design and digital verification. They can be used to create an "AI linter" for digital designs and testbenches.

There are 10 fields in each row:
* `title`: a very short human-readable description of the guideline. Useful mostly for uniquifying the guidelines.
* `statement`: a single sentence of what a user should do to follow the guideline
* `description`: a paragraph of text going into further detail about what the guideline is
* `scope`: which part of a digital design/verification testbench a design would apply to
* `not-scope`: is like the semantic opposite of `scope`, in the sence that the guideline does not apply to this
* `justification`: why the guideline should be followed; what might happen if the guideline isn't followed
* `exceptions`: common exceptions to the guideline. If there are no exceptions, then the text `No exceptions.` or `No exceptions`.
* `consequences`: a list of strings with possible consequences of not following the guideline/rule. Can be used to derive a hazard rating.
* `tag_0`: either `verification` or `design`
* `tag_1`: either `architecture`, `requirements`, `specification`, or `implementation`

## Data Source
We created the dataset by using DeepSeek V3 as a direct data generator. We asked the model to produce examples of guidelines that matched what we were looking for. By adjusting the temperature of the model and feeding back already-generated examples back to the model, we were able to generate thousands (~21K) of guidelines without relying on external sources.

After collecting the outputs, we refined them through automated filtering pipeline and semantic clustering. Low-quality or off-topic samples were removed, and related examples were grouped together so we could keep only the most consistent and useful ones.

## Credits
Gilberto Lopez Astorga, Richard Tilkov, and Vlad Pomogaev