README.md

---
language_creators:
- expert-generated
language:
- code
license:
- other
multilinguality:
- multilingual
pretty_name: Luau-Stack-HQ
task_categories:
- text-generation
tags:
- luau
- code
- roblox
- corpus
size_categories:
- 10K<n<100K
---

# Luau Stack HQ

Luau Stack HQ is a highly curated & deduplicated Luau source code from github repository with linter `(stylua)`, it contains **~16.29M** tokens `(21515 unique files & 259 unique repo)`. It is built strictly for the training stages, providing a pristine density of type-safe, functional architecture without the noise of standard web scrapes.

## # Filtering Pipeline

To ensure the highest quality data for pretraining, the dataset was built using a rigorous, multi-step extraction and filtering pipeline:

* **Repository Selection:** * **Target Terms:** Searches focus on `rbx`, `rblx`, `roblox`, and `luau`.
    * **Quality:** Minimum of 5 stars.
    * **Relevance:** Because Luau is updated frequently, only repositories active or pushed within the last 2.5 years are included.
    * **Repo Deduplication:** Forks are strictly excluded.
* **Auto-Formatting (`StyLua`):** Every raw `.luau` and `.lua` file is piped through `StyLua` to ensure consistent syntax. Unparseable files are dropped.
* **Cheat & Exploit Removal:** * Repositories with cheat-related keywords in their name, description, or topics (e.g., `aimbot`, `executor`, `synapse`, `krnl`) are blacklisted.
    * Code containing Roblox exploit globals (e.g., `getfenv`, `setfenv`, `hookfunction`, `loadstring`, `getrawmetatable`) is entirely discarded.
* **Noise Reduction:** Test files, mocks, `.spec.` files, and compiled `/dist/` directories are ignored. Files smaller than 100 bytes or larger than 1MB are also excluded.

## # Licensing

The source code contained within this dataset belongs to their respective original authors. All data was sourced from public GitHub repositories under open-source licenses, with forks explicitly excluded. Please respect the original repository licenses when utilizing this dataset for commercial model training.