Hugging Face Video

This page collects nine videos about Hugging Face, from short "what is it" explainers to hands-on sessions that run models in Google Colab, use KerasHub with Hub checkpoints, and introduce the Hugging Face AI Agents course. Below the player grid there is a suggested viewing order, a summary of the concepts the videos keep returning to, a minimal code example to try alongside them, and the common mistakes beginners make when moving from a notebook demo to their own project.

1. Suggested learning path

The videos overlap, so watching all nine in grid order repeats some material. Begin with one of the two "What Is Hugging Face" explainers and "Hugging Face and How To Use It" to learn the vocabulary: models, datasets, Spaces, and the model card. Then follow one of the Colab sessions while typing along, because loading and running a model yourself is where the concepts click. After that, pick a direction: KerasHub and the Bootcamp go deeper on libraries and fine-tuning, while the AI Agents course and the HuggingCast livestream look at tool-using agents built on top of Hub models.

2. Core concepts the videos cover

Three ideas recur across the videos. First, everything on the Hub is a git repository identified as owner/name, with versions you can pin. Second, the transformers pipeline wraps tokenizer, model, and post-processing behind one function call for common tasks such as sentiment analysis, summarization, or text generation. Third, the Auto classes (AutoTokenizer, AutoModelFor...) read a repository's configuration and instantiate the right architecture, which is how one line of code can load thousands of different models.

3. The videos

Click a thumbnail to play it in place, or use the Play All button to queue all nine. Each video opens inside this page, so you can keep the code example below open in another tab while you watch.

Hugging Face and How To Use It

Running HuggingFace Models in Google Colab

What Is Hugging Face

Hugging Face's AI Agents Course

Running HuggingFace Models in Google Colab

What Is Hugging Face

How to use KerasHub with Hugging Face

Hugging Face Bootcamp

HuggingCast v1

4. Try it: the first pipeline

This is the example most of the hands-on videos start with. It runs on CPU, downloads a small default model on first use, and caches it under ~/.cache/huggingface.

# pip install transformers torch
from transformers import pipeline

classifier = pipeline("sentiment-analysis",
                      model="distilbert/distilbert-base-uncased-finetuned-sst-2-english")
print(classifier(["I love how easy this is.", "The download took forever."]))
# [{'label': 'POSITIVE', 'score': ...}, {'label': 'NEGATIVE', 'score': ...}]

Naming the model explicitly, rather than relying on the pipeline's default, keeps results reproducible when library defaults change. In Colab, switch the runtime to a GPU and pass device=0 to the pipeline for larger models.

From pipeline to explicit model. Once the pipeline works, the next step the longer videos take is to load the tokenizer and model directly with AutoTokenizer.from_pretrained and AutoModelForSequenceClassification.from_pretrained, tokenize a batch, run a forward pass, and apply softmax to the logits. Doing it by hand shows what the pipeline hides: padding and truncation, the maximum sequence length, label mappings in the model configuration, and where tensors live (CPU or GPU). That understanding is what you need before fine-tuning with the Trainer API or exporting a model for production serving.

5. Pitfalls when moving beyond the demos

Memory. A model's parameter count times bytes per parameter (2 for fp16, about 0.5 for 4-bit) is a lower bound on memory; demos in Colab often use small models for this reason.
Gated models. Llama and some image models require accepting a license and logging in with a token before download.
Licenses. Check the model card's license before commercial use; "open weights" does not always mean unrestricted.
Unpinned revisions. Repositories can be updated; pin a commit for anything you deploy.
Untrusted code. trust_remote_code=True runs Python from the repository; only enable it for sources you trust.

Hugging Face control center
Hugging Face Hub and Datasets
Transformers
NVIDIA + CUDA setup for local GPU inference