llm-text-correct

This skill mostly does what it says (KenLM + MacBERT correction) but has several surprising behaviors you should consider before installing: - It will create a virtual environment and install large packages (PyTorch, audio libs, huggingface-hub) and download large models (~400MB+). Expect GBs of downloads and long install times; do this only on machines where that is acceptable. - If you point it at a folder, it will recursively edit many file types including .py, .json, .js and other code/config files — the README/SKILL.md doesn't clearly warn about this. Back up any folders before running and test on a small sample first. - The scripts install unrelated audio packages (audio-separator, librosa, pydub) which are not needed for text correction; that suggests borrowed code and increases the attack surface. Consider removing or auditing those lines if you only need text correction. - It will write a venv, models, and logs under the skill root (or nearby directories). If you prefer isolation, run the code in an isolated VM or container, or inspect/modify the scripts to change VENV_DIR/MODELS_DIR paths. - The code uses legit external services (PyPI mirror, download.pytorch.org, huggingface.co) rather than obfuscated endpoints, but automatic network access occurs on first run. If you need to control downloads, use --model-path to provide models locally and run in an environment with no network. Recommended actions: 1) Review scripts locally and run them on a copy of your data. 2) Run in an isolated environment (container/VM) if you are unsure. 3) Remove or modify audio-related installs and the get_text_files extension list if you want to avoid touching code files. 4) Ensure you have sufficient disk and bandwidth for model downloads.