download/load the HF dataset

download_dataset is the first step in the LLM School pack's training pipeline: it grabs a dataset from Hugging Face and saves it to disk, so the next node (split_dataset) can work with local files instead of re-fetching over the network. It's a thin wrapper around huggingface_hub.snapshot_download() with repo_type="dataset", marked as an output node so it sits at the end of a data-preprocessing chain and visibly reports its result.

You feed it a repo_id - the owner/name handle of any public dataset on the Hub - and it downloads the whole thing into a cache folder inside the pack (<pack>/datasets, or wherever you point cache_dir). The output is a single string: the absolute local path to the downloaded files. That's the local_dir you'll paste into split_dataset.

The inputs that matter

• repo_id - default rajpurkar/squad_v2. This is the dataset the pack's trainer was written against (SQuAD-style question/context/answers), so it's the safe starting point. Swap in anything you found with get_dataset_name.
• cache_dir - default datasets, resolved relative to the pack folder. Relative paths can surprise you; if you want the data somewhere predictable, give an absolute path.
• token - the trap. The default is the literal string hf_XXX. Leave it there and the code passes hf_XXX as your auth token, which fails with a 401 on anything that actually requires auth. Set it to a real token from https://huggingface.co/settings/tokens, or set it to an empty string so it falls back to the HF_token value in the pack's config.ini. Public datasets like squad_v2 don't need a token at all.
• force_download - defaults to True, which means it re-downloads everything on every run. Turn it off once you've got a copy locally; it's the default that'll waste your bandwidth for no reason.

How to install

ComfyUI Manager, search "comfyui_LLM_schools" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/heshengtao/comfyui_LLM_schools

then restart ComfyUI. requirements.txt is light - huggingface_hub, datasets, transformers, peft - and Manager installs it for you.

Common issues

• The node doesn't appear at all. The pack reads config.ini at import time, and if that file doesn't exist the whole dataset.py module fails to load (the pack's __init__.py swallows the error). The repo ships config.ini.example; copy it to config.ini inside the pack folder before you restart, then put a real token (or #zh_CN/en_US for the language) in it.
• "The default token isn't accepted." See above - hf_XXX is a placeholder, not a valid token. Empty the field for public datasets.
• Deprecation warnings or an error about local_dir_use_symlinks. The code passes that argument to snapshot_download, and newer huggingface_hub versions deprecated/removed it. Usually just a warning; if it hard-fails, update huggingface_hub to a current version where the parameter is tolerated or ignored.

It's a simple, opinionated node - the opinion being "download SQuAD-style QA datasets and fine-tune causal LMs on them." If that's your plan, this is the cleanest way to get data onto disk from inside ComfyUI.