Cheatsheet: Share your work on the Hub

Authenticate first

from huggingface_hub import notebook_login
notebook_login()      # notebook
# hf auth login   # terminal

Paste a token from your account settings; it is cached for later uploads.

The three routes (easiest to most manual)

Route	Use when	Key calls
push_to_hub API	Almost always	push_to_hub=True, trainer.push_to_hub(), model.push_to_hub()
huggingface_hub library	Programmatic repo management	create_repo, upload_file
git + git-lfs	Full manual control	git lfs install, clone, add, commit, push

Route 1: push_to_hub (preferred)

# During Trainer training:
training_args = TrainingArguments(
    "bert-finetuned-mrpc", save_strategy="epoch", push_to_hub=True
)
# ... trainer.train() ...
trainer.push_to_hub()      # final push + auto-generated model card

# Or directly on objects:
model.push_to_hub("dummy-model")
tokenizer.push_to_hub("dummy-model")     # always push the tokenizer too

hub_model_id="org/name" targets an organization. trainer.push_to_hub() auto-writes a card with hyperparameters + metrics.

Route 2: huggingface_hub library

from huggingface_hub import create_repo, upload_file
create_repo("dummy-model")                 # private=True to hide; repo_type="dataset"/"space"
upload_file("config.json", path_in_repo="config.json",
            repo_id="username/dummy-model")   # HTTP, no git; <5GB per file

Route 3: git + git-lfs (barebones)

git lfs install
git clone https://huggingface.co/username/dummy-model
# model.save_pretrained(dir) + tokenizer.save_pretrained(dir)
git add . && git commit -m "First model version" && git push

git-lfs auto-tracks large weight files; small JSON files use plain git.

What a usable model repo contains

File	What it is	Handler
config.json	Architecture + settings	git
weights (*.bin / *.safetensors)	The trained parameters	git-lfs
tokenizer files	Vocab, tokenizer_config.json, special tokens	git
README.md	The model card	git

This is exactly the set from_pretrained expects, which is why one line loads it.

The model card (README.md) essentials

• What it is for (the task, the intended use)
• Where it came from (base checkpoint, training data)
• What it cannot do (limitations, biases, language scope)
• How to use it (a short code snippet)

A good card is the difference between a model others adopt and weights nobody touches.

Words to use precisely

• Namespace: your username or an organization you can write to.
• git-lfs: Git Large File Storage; tracks large binaries (weights) separately from plain git.
• Model card: the repo’s README.md; documentation + honesty statement.

Recommended further study

• Hugging Face LLM Course, Chapter 4: “Sharing models and tokenizers.” huggingface.co/learn/llm-course/chapter4. Released under Apache 2.0; this lesson mirrors its structure with original prose.