Browsing, Filtering, Naming Conventions

The filter axes are the mental model

The /models browse page exposes the same filter set the API does: task (pipeline_tag), library (transformers, diffusers, sentence-transformers, mlx, peft, ...), language (ISO 639-1), license, provider (which inference providers serve this model), size, and format. Sorting axes: trending, downloads, likes, recently updated, recently created. Treating these as the actual taxonomy of the Hub — not the homepage chrome — will save you weeks of "I can't find a model that does X."

Naming is a contract, not a vibe

HF model IDs follow {org}/{base}-{size}-{variant} conventions. A few patterns to internalize:

meta-llama/Llama-3.1-8B — base model. No instruction tuning.
meta-llama/Llama-3.1-8B-Instruct — instruction-tuned variant. Use this for chat unless you specifically want raw base behavior.
TheBloke/Llama-2-7B-Chat-GGUF — quantized GGUF for llama.cpp / Ollama.
{org}/{model}-AWQ / -GPTQ / -bnb-4bit — quantized variants for inference servers.
{org}/{model}-Matryoshka / -Distill — specialized derivatives (embedding truncation, distillation).

The naming is informational, not enforced. base_model and base_model_relation in the YAML front-matter (next lesson) are the machine-readable version of the same lineage.

Code

Filter on multiple axes·python

from huggingface_hub import HfApi

api = HfApi()

# Korean text-generation models, Apache-2 only, Transformers-loadable
results = api.list_models(
    task="text-generation",
    language="ko",
    library="transformers",
    sort="downloads",
    limit=10,
)
for m in results:
    print(f"{m.id:<55} dl={m.downloads or 0:>8}  {','.join((m.tags or [])[:4])}")

Full-text search across cards·python

from huggingface_hub import HfApi
api = HfApi()

# Searches model card / readme content, not just the id
results = api.list_models(
    search="korean instruction tuning",
    sort="downloads",
    limit=5,
)
for m in results:
    print(m.id)

Exercise

Open huggingface.co/models. Apply filters until you have at most 20 candidates: task = text-generation, language = en (or ko), library = transformers, license = apache-2.0 or mit, parameters between 1B and 8B. Sort by trending. From the result, pick three you'd genuinely consider for an on-prem project. Note the differences in their model cards (next lesson) for each.