ranjithraj/hf-landscape-study-data

🗺️ HF Landscape Study Data

Parquet crawl data powering HF Landscape — a leaderboard and ecosystem stats dashboard for the Hugging Face Hub.

What's Inside

Six Parquet files covering the full Hub at crawl time, generated from a DuckDB database:

File	Records	Description
models.parquet	~2.9M	Every model: downloads (30d + all-time), likes, task, library, params, license, language, country, entity type, trending score, modality, size bucket
datasets.parquet	~955K	Every dataset: downloads, likes, trending score, task categories, license, country, entity type
spaces.parquet	~1.4M	Every space: likes, trending score, SDK, country, entity type
collections.parquet	~217K	Every collection: upvotes, item count, country, entity type
entities.parquet	~1.4M	Per-entity aggregation: one row per author with rolled-up counts across all repo types
entities_all.parquet	~5.5M	All repos (models + datasets + spaces + collections) in one file, with a repo_type field to filter by type

Schema

models.parquet

Field	Type	Description
id	string	Repo ID (org/name)
author	string	Organization or user
dl30	double	Rolling 30-day downloads
dlAll	double	All-time downloads
likes	double	Total likes
task	string?	Pipeline tag (e.g. text-generation)
params	double?	Parameter count (from safetensors metadata)
langs	list<string>	Language tags
createdAt	timestamp	Creation timestamp
license	string?	License identifier
baseModel	string?	Base model name
baseRelation	string?	Relation to base (quantized, adapter, finetune, …)
library	string?	Framework (e.g. transformers)
lastModified	timestamp	Last commit timestamp
gated	string?	Whether the repo is gated
trending	double	Trending score
modality	string?	Input modality (nlp, cv, multimodal, audio, …)
sizeBucket	string?	Parameter count bucket (<5M, 5M–100M, 100M–500M, 0.5B–1B, 1B–5B, 5B–15B, 15B–70B, 70B+, unknown)
country	string	Country code from hand-annotated entity map (- = unmapped)
entityType	string	company, community, individual, or unknown

datasets.parquet

Field	Type	Description
id	string	Repo ID
author	string	Owner
dl30	double	Rolling 30-day downloads
dlAll	double	All-time downloads
likes	double	Total likes
trending	double	Trending score
taskCategories	list<string>	Task categories
license	string?	License identifier
createdAt	timestamp	Creation timestamp
country	string	Country code
entityType	string	Entity type

spaces.parquet

Field	Type	Description
id	string	Repo ID
author	string	Owner
likes	double	Total likes
trending	double	Trending score
sdk	string?	Space SDK (docker, static, gradio, streamlit, …)
createdAt	timestamp	Creation timestamp
country	string	Country code
entityType	string	Entity type

collections.parquet

Field	Type	Description
slug	string	Collection slug
owner	string	Owner username
title	string	Collection title
upvotes	double	Total upvotes
itemCount	double	Number of items
lastUpdated	timestamp	Last update timestamp
country	string	Country code
entityType	string	Entity type

entities.parquet

Per-entity aggregation — one row per author with rolled-up stats across all repo types.

Field	Type	Description
name	string	Author / org name
country	string	Country code
type	string	Entity type
models	int	Number of models
datasets	int	Number of datasets
spaces	int	Number of spaces
collections	int	Number of collections
downloads_all_time	double	Total all-time downloads across all repos
downloads_30d	double	Total 30-day downloads across all repos
likes	double	Total likes across all repos
upvotes	double	Total collection upvotes

entities_all.parquet

All repos in one file. Each record carries the full fields of its source type plus a repo_type discriminator.

Field	Type	Description
repo_type	string	model, dataset, space, or collection
(+ all fields from the matching source schema above)

Country Attribution

Country and entity type are from a hand-maintained map (data/spotlight.json) applied at crawl time. The Hub API exposes no location field, so attribution is inferred from public signals — not a verified fact. Only ~1,200 mapped entities carry a country; the rest are - (unknown).

Usage

With Polars (recommended):

import polars as pl

df = pl.read_parquet("models.parquet")
top = df.sort("dl30", descending=True).head(10).select("id", "dl30", "dlAll", "likes")
print(top)

Filter entities by type:

import polars as pl

df = pl.read_parquet("entities_all.parquet")
datasets = df.filter(pl.col("repo_type") == "dataset")
print(datasets.sort("dlAll", descending=True).head(10).select("id", "author", "dlAll"))

Query list columns (e.g. find models with a specific language):

import polars as pl

df = pl.read_parquet("models.parquet")
hindi = df.filter(pl.col("langs").list.contains("en"))
print(f"English models: {len(hindi):,}")

With DuckDB:

import duckdb

con = duckdb.connect()
df = con.execute("SELECT * FROM 'models.parquet' WHERE task = 'text-generation' LIMIT 10").fetchdf()
print(df)

Regeneration

Data is crawled weekly from the Hugging Face Hub API and exported as Parquet via the HF Landscape CLI:

npm link          # install hf-study CLI
hf-study          # full crawl + aggregation
npm run db:build  # export Parquet files

Citation

@dataset{hf_landscape_2026,
  title = {HF Landscape Study Data},
  author = {Ranjith Raj},
  year = {2026},
  url = {https://huggingface.co/datasets/ranjithraj/hf-landscape-study-data}
}