huggingface-model-search

HuggingFace Model Search

Find a pretrained model on the HuggingFace Hub to fine-tune or use zero-shot. Use this when the task spec names a dataset/metric and you need a strong backbone — e.g. AIRS-Bench tasks where training from scratch is off the table.

When to use

• The task asks you to beat a SOTA paper on a named dataset — find the model the paper (or a close descendant) released on HF.
• You need a backbone of a particular size (e.g. ~7B, ≤16B) for a task family (text classification, QA, code gen, molecular property, time series, ...).
• You want to check whether a candidate model actually exists + has the expected files (config, tokenizer, weights) before writing training code.

Method 1 — HF Hub API (preferred)

huggingface_hub is already available in most workspaces (or installable via uv add huggingface_hub). Use HfApi().list_models(...) — it supports filters for task, library, language, and sorts by downloads/likes/trending.

from huggingface_hub import HfApi
api = HfApi()

# Example: top-downloaded text-classification models, sorted
models = api.list_models(
    task="text-classification",           # pipeline_tag filter
    sort="downloads",                     # or "likes", "trending", "lastModified"
    direction=-1,
    limit=30,
    search="sentiment",                   # free-text name/tag search (optional)
)
for m in models:
    print(m.id, m.downloads, m.tags[:6])

Common task= values: text-classification, text-generation, question-answering, token-classification, translation, summarization, sentence-similarity, image-classification, time-series-forecasting, graph-ml. The full list is at https://huggingface.co/tasks.

Filter by size (rough — use tags + model card)

list_models doesn't expose a numeric param count, so filter by the tags convention (7b, 13b, mistral, llama) or by the org/name prefix, then confirm by reading the model card:

models = api.list_models(task="text-generation", tags=["7b"], sort="downloads", limit=20)