ai4science-huggingface-recipes

name	ai4science-huggingface-recipes
description	Guides creation and editing of Hugging Face–centric recipes in AI4Science Studio—discovery, task alignment, scripting, documentation, and attribution.

Hugging Face recipes workflow

Use this skill when adding or refactoring train / fine-tune / inference / eval material for models hosted on Hugging Face inside this repo.

Steps

1. Identify the model on Hugging Face (model card, license, intended use, dependencies). Record the full id org/model in the model folder README.md.
2. Map to a folder name under the correct domain’s models/ directory. Default: replace / with __ in the Hub id (org__model). Alternative: a public model name on disk (e.g. HydraGNN, ORBIT-2) is acceptable when that domain’s models/README.md allows it and the model README.md states the full org/model id at the top—do not rename those folders to org__model unless the user asks.
3. Align the task with what the model card and upstream code support (inference-only vs fine-tuning vs training from scratch).
4. Reuse upstream scripts or minimal wrappers: prefer calling official examples with pinned versions rather than reimplementing full training stacks unless necessary.
5. Document Python/PyTorch (or other) versions, ROCm builds and versions for AMD Instinct when validated, and how to run from the repo root or from recipes/ subfolders.
6. Attribute authors, papers, and license in the model README.md and in recipe comments where helpful.

Recipe folder tips

• Use subfolders under recipes/ for distinct tasks, e.g. recipes/inference/, recipes/finetune/.
• Keep entrypoints small and documented; link to Hugging Face Spaces or Colab only as supplements, not replacements for reproducible commands.

Avoid

• Committing token values or private Hub tokens.
• Checking in large *.bin, *.safetensors, or full datasets when .gitignore already excludes them—point users to Hub or documented download steps instead.
• Hardcoding HF repo filenames without verifying: model repos ship different names than recipes assume (e.g. walrus.pt not model.pt). Use list_repo_files() to discover actual names, filter by extension, pick smallest or first match.
• Assuming a _mi300x.sh SLURM script name — use _amd.sh (covers MI250X, MI300X, MI350X with the same rocm7.2.x image).

HF checkpoint download pattern (robust)

from huggingface_hub import list_repo_files, hf_hub_download
import glob, os

repo = "org/model"
cache = os.path.join(os.environ.get("HF_HOME", os.path.expanduser("~/.cache/huggingface")), "model")
os.makedirs(cache, exist_ok=True)

# Reuse any existing download
existing = glob.glob(os.path.join(cache, "**/*.ckpt"), recursive=True)
if existing:
    chosen = sorted(existing)[0]
else:
    files = [f for f in list_repo_files(repo) if f.endswith(".ckpt")]
    chosen_name = sorted(files)[0]  # or filter by size keyword
    chosen = hf_hub_download(repo_id=repo, filename=chosen_name, local_dir=cache)
print(chosen)