AI Intermediate: Hugging Face & Open Source Models Practice Questions

pipeline() is correct because it provides a high-level, task-oriented API that automatically handles model loading, tokenization, inference, and post-processing in a single call — ideal for quick testing without boilerplate code. AutoModel.from_pretrained() loads only the model weights and requires manual tokenization and forward-pass logic. AutoTokenizer.from_pretrained() loads only the tokenizer, not the model. Trainer() is used for fine-tuning workflows, not inference.

Apache 2.0 and MIT are correct because both are permissive open-source licenses that allow unrestricted commercial use, redistribution, and modification with minimal conditions (attribution only). The Llama Community License restricts use based on monthly active user thresholds (>700M MAU requires a separate Meta agreement) and contains acceptable-use provisions on outputs. RAIL imposes behavioral-use restrictions on what the model's outputs can be used for, making it unsuitable here. CC BY-NC 4.0 explicitly prohibits commercial use (NC = NonCommercial).

load_in_4bit=True in the from_pretrained() call is a BitsAndBytes integration feature. The bitsandbytes library provides the NF4 4-bit quantization kernels that the transformers library wraps via this parameter. optimum is a Hugging Face library for hardware-specific optimizations (ONNX, TensorRT, etc.) but does not provide 4-bit loading. auto-gptq handles GPTQ-quantized models and uses a different loading pattern. llama-cpp-python is a Python binding for llama.cpp and works with GGUF files, not the standard from_pretrained() API.

A Hugging Face Space running a Gradio app is correct. Spaces are purpose-built for hosting interactive ML demos with zero infrastructure management; they auto-provision compute, provide a public URL, and support Gradio (and Streamlit) apps natively. The Inference API provides programmatic model inference but not an interactive web UI for stakeholders. A model card hosts documentation and metadata, not an interactive app. Trainer with push_to_hub=True uploads model weights to the Hub but does not create a runnable demo.

In an Ollama Modelfile, SYSTEM sets the system prompt (e.g., 'You are a pirate'), and PARAMETER num_ctx 2048 sets the context window size. Together they achieve both goals. PROMPT and CONTEXT are not valid Modelfile instructions. TEMPLATE controls the full prompt template format (not the system persona), and LIMIT does not exist. INSTRUCTION is not a Modelfile directive; the base model is specified with FROM.