Does loading this Hugging Face repo run code? Loadcheck is a static, browser-only checker for a small set of observable loading-risk signals in public Hugging Face model and dataset repositories.
Not maintained since 2026-08-15.
Requires Node 22 or newer.
npm run build
npm testServe dist/ with any static file server. The browser contacts only the same origin and https://huggingface.co.
Loadcheck resolves the submitted public repository to a commit SHA before reading its tree. It reads only config.json for models and README.md front matter for datasets, with a 1 MiB limit per file. It does not download weights, execute repository code, render templates, or use a Hub token.
It reports custom auto_map mappings that current Transformers loads execute only after an explicit trust_remote_code=True opt-in; legacy Datasets v3.4 root loading scripts, whose repository code ran only after the same explicit opt-in; and pickle-capable model paths when no .safetensors path exists. Template-like delimiters below dataset README configs are neutral manual-review signals, not evidence of code execution or a pinning requirement: current YAML builder parameters work without custom code, and no current standard Hugging Face loading call is documented to render those delimiters. API, parsing, size, pagination, rate-limit, and browser network failures fail closed as incomplete checks.
The submitted repository ID goes from the browser to Hugging Face. Loadcheck sends one fixed-path, same-origin completion request after a terminal result; it has no body, query string, repository ID, type, revision, result, or timestamp. There are no cookies, local storage, service worker, analytics, or backend.
- Transformers supported models
- Transformers Auto classes and
trust_remote_code - Legacy Datasets v3.4 loading scripts
- Hub pickle scanning
- Dataset manual configuration
MIT. See LICENSE.
Questions: hello@cybernative.ai