Every model card on Hugging Face is a README.md file. At the very top, between two lines of ---, sits a block of YAML (Yet Another Markup Language). This is the structured metadata that Hugging Face parses automatically. Everything below the second --- is free-form Markdown for humans. The YAML block is what powers search, filtering, the inference widget, and the sidebar badges you see on every model page.

This field determines whether you can legally use the model. Common values: apache-2.0 (fully permissive, commercial OK), mit (permissive), llama3.1 (Meta’s community license with usage restrictions above 700M monthly users), gemma (Google’s terms with prohibited use cases), cc-by-nc-4.0 (non-commercial only). If you’re building a product, this is the first field to check. A model with cc-by-nc-4.0 cannot be used commercially, period.

Uses ISO 639-1 codes: en for English, zh for Chinese, fr for French. A model listing [en, de, fr] was trained on those languages. A model listing only [en] may produce garbage output in other languages, even if it technically generates text. Multilingual models typically list many codes or use multilingual as a tag.

This tells you the model’s primary task. Common values for LLMs: text-generation (autoregressive generation), text2text-generation (encoder-decoder like T5), fill-mask (BERT-style masked language modeling). For other modalities: text-to-image (Stable Diffusion), automatic-speech-recognition (Whisper), image-classification, feature-extraction (embedding models). The pipeline tag also powers the interactive widget on the model page — it tells HF what kind of input box to show.

How to load the model in code. transformers (Hugging Face’s main library), diffusers (for diffusion models), sentence-transformers (for embeddings), peft (for adapters/LoRA), gguf (for llama.cpp format). This tells you which import to use: a transformers model loads with AutoModelForCausalLM.from_pretrained(), while a gguf model loads with llama.cpp or Ollama.

The base_model field links to the parent model this one was derived from. A fine-tuned model will point to its foundation model; a quantized variant will point to the full-precision original. Example: base_model: meta-llama/Llama-3.1-8B tells you this is a derivative of Meta’s Llama 3.1 8B. You can click through to the base model to see its original card, benchmarks, and training data.

Models often form a chain: Base → Fine-tune → Quantized. For example: meta-llama/Llama-3.1-8B (base) → NousResearch/Hermes-3-Llama-3.1-8B (fine-tune) → bartowski/Hermes-3-Llama-3.1-8B-GGUF (quantized). Each link in the chain inherits the upstream model’s strengths, weaknesses, and license terms. A fine-tune of a Llama model still carries the Llama Community License, regardless of what license the fine-tuner claims.

Lists the Hugging Face dataset IDs used for training. Example: datasets: [HuggingFaceTB/cosmopedia, allenai/dolma]. This lets you click through to the actual training data and inspect it. For fine-tuned models, this usually lists the fine-tuning dataset, not the base model’s pre-training data. Watch for models that don’t list any datasets — either the data is proprietary or the model creator didn’t document it.

Free-form labels that help with discovery. Common useful tags: chat (instruction-tuned for conversation), code (trained on code), math (math-focused), gguf, 4bit, lora. Tags are not validated — anyone can add any tag. They’re useful for broad filtering but should not be trusted as ground truth. Cross-check tags against the actual card content.

The model-index field embeds benchmark results directly in the metadata. Hugging Face uses a decentralized evaluation system: results are stored as YAML files in an .eval_results/ folder in the model repo. These results appear automatically on the model page with badges showing their provenance — “verified” (ran on HF infrastructure), “community” (submitted via PR), or “leaderboard” (from the Open LLM Leaderboard).

Check the badge type: verified results are more trustworthy than self-reported ones. Check the benchmark names: do they cover the tasks you care about? And check whether the results are suspiciously high — if a 7B model claims 95% on MMLU, be skeptical. The evaluation results also link to the benchmark dataset’s leaderboard, letting you compare this model against others on the same benchmark.

When you go to huggingface.co/models and use the filters on the left sidebar, every filter maps directly to a YAML field. Filter by “Text Generation”? That’s pipeline_tag. Filter by “English”? That’s language. Filter by “Apache 2.0”? That’s license. Filter by “transformers”? That’s library_name. A model with no YAML metadata is a model that doesn’t appear in any filtered search.

The interactive widget on the model page (the text box where you can type a prompt and see output) is powered by pipeline_tag. If the tag says text-generation, you get a text input. If it says text-to-image, you get an image generation interface. If it says automatic-speech-recognition, you get a file upload for audio. No pipeline_tag = no widget = no way to test the model in-browser before downloading.

When you land on a model page, read the YAML in this order:

1. license — Can I use this? (If non-commercial and you need commercial, stop.)
2. pipeline_tag — Is this the right task type?
3. language — Does it support my language?
4. base_model — What family does it belong to?
5. library_name — Can I load it with my stack?
6. datasets — What was it trained on?
7. tags — Any useful context?

If all 7 checks pass, you have a candidate worth investigating. Now you move beyond the YAML: read the prose description, check the benchmark tables (Chapter 4), look at the Files tab (Chapter 6), and scan the Community discussions. If any of the 7 checks fail, move on to the next model — there are over 2 million models on the Hub. Your time is better spent finding a model that fits than trying to make a misfit work.