Hugging Face & Open Source Models
This domain tests your ability to work with the Hugging Face ecosystem—Hub, APIs, and the transformers library—as well as open-source model selection, local inference with Ollama, quantization strategies, and fine-tuning fundamentals. It accounts for 17% of the exam, making it one of the heavier-weighted domains. Expect questions on exact API behavior, license restrictions, quantization tradeoffs, and when to choose one tool or approach over another.
Hugging Face Hub: Model Cards, Datasets, Spaces, Organizations
The Hugging Face Hub is a Git-based repository platform hosting models, datasets, and Spaces. Every artifact lives at a URL with the pattern huggingface.co/{owner}/{repo-name}.
Model cards are README.md files stored at the root of a model repository. They contain: intended use, training data, evaluation results, license, and limitations. The Hub renders them as a structured page. Exam questions may ask what component documents bias, limitations, or usage restrictions—the answer is always the model card.
Datasets on the Hub are versioned repositories loadable via datasets.load_dataset("owner/name"). They support multiple splits (train/validation/test) and multiple configurations (subsets).
Spaces host live demos backed by Gradio or Streamlit. Free Spaces run on CPU; GPU Spaces require a paid tier or grant. Spaces use a README.md with YAML front matter that sets sdk: gradio or sdk: streamlit.
Organizations let teams share repositories under a single namespace. Members have roles: read, write, contributor, admin.
| Hub Object | Storage Format | Key File | Loaded With |
|---|---|---|---|
| Model | Git LFS repo | config.json, pytorch_model.bin / safetensors | from_pretrained() |
| Dataset | Git LFS repo | dataset_infos.json, Parquet/CSV/JSON | load_dataset() |
| Space | Git repo | app.py, README.md (YAML header) | Browser / iFrame embed |
Common traps:
- Confusing the model card (documentation) with
config.json(architecture config). The model card isREADME.md. - Assuming Spaces always have GPU—free tier is CPU-only.
- Thinking
load_datasetrequires a local file; it can pull directly from the Hub by name.
Model Licenses: Apache 2.0, MIT, Llama Community License, RAIL
License choice determines whether you can use a model commercially, redistribute it, or modify it.
| License | Commercial Use | Redistribution | Modification | Key Restriction |
|---|---|---|---|---|
| MIT | ✅ Yes | ✅ Yes | ✅ Yes | Must include copyright notice |
| Apache 2.0 | ✅ Yes | ✅ Yes | ✅ Yes | Must state changes; patent grant included |
| Llama Community License | ✅ Yes (≥700M MAU requires Meta approval) | ✅ Yes (must include license) | ✅ Yes | Cannot use outputs to train competing foundation models |
| RAIL (Responsible AI License) | ⚠️ Conditional | ✅ Yes | ✅ Yes | Prohibits specific harmful use cases enumerated in the license |
| OpenRAIL-M | ✅ Yes | ✅ Yes | ✅ Yes | Same use-case restrictions as RAIL; common on HF |
Key distinctions for the exam:
- Apache 2.0 provides an explicit patent license—MIT does not. If patent protection matters, Apache 2.0 is the stronger choice.
- Llama Community License blocks using model outputs to train competing foundation models, and organizations with >700M monthly active users must request a separate license from Meta.
- RAIL/OpenRAIL-M licenses travel with the model—downstream users must also comply with use restrictions. They do NOT prohibit commercial use broadly; they prohibit specific enumerated harms.
Common traps:
- Thinking Llama models cannot be used commercially—they can, with the exception noted above.
- Confusing RAIL as "open source"—the OSI does not recognize RAIL as open source because of use restrictions.
- Assuming MIT and Apache 2.0 are interchangeable—Apache 2.0 adds patent protection and change documentation requirements.
Worked scenario: A startup wants to fine-tune Llama 3.1 8B for a customer-service product serving 500M users and redistribute the fine-tuned model. Is this allowed? Yes—500M < 700M MAU threshold. If they grow past 700M MAU, they must contact Meta for a license. Using those fine-tuned model outputs to train a new competing LLM is prohibited regardless of size.
Unlock with a paid plan
11 more sections (about 2,684 words) continue below. Every paid plan unlocks all AI Intermediate study guides, the full question bank and mock exams, from $4.97/month.