← Blog

verba on Hugging Face: Load and Use the Corpus

2026-06-24

The verba corpus is now available on Hugging Face, the largest hub for machine-learning and NLP datasets. That means you can pull all 48,787 proverbs and start working with them in a single line of code — no sign-up, no API keys, no manual file downloads.

This article covers four ways to get the corpus (the datasets library, pandas, the hf CLI, and SQL via DuckDB), plus a few examples of what to do with it next.

What's published

Dataset page: huggingface.co/datasets/dmytro-yemelianov/verba.

The repository holds four files:

• corpus.csv — the canonical corpus: 48,787 records, 10 columns. This is the train split that the dataset viewer renders and the datasets library loads.
• sources.csv — the registry of the five source collections (bibliography).
• croissant.json — MLCommons Croissant metadata for ML pipelines.
• README.md — the full data card (methodology, quality limits, licensing).

The dataset is public and licensed CC BY 4.0 (compilation and enrichment); the historical texts are in the public domain. No token is required to read it.

1. The datasets library (Python)

The simplest path for researchers and ML engineers:

from datasets import load_dataset

ds = load_dataset("dmytro-yemelianov/verba", split="train")
print(len(ds))                       # 48787
print(ds[0]["text"], "→", ds[0]["modern_text"])

# keep only proverbs of a given theme
work = ds.filter(lambda r: "work_labor" in (r["category"] or ""))
print(len(work))

For large workloads or limited memory, turn on streaming — nothing is written to disk:

ds = load_dataset("dmytro-yemelianov/verba", split="train", streaming=True)
for row in ds.take(5):
    print(row["modern_text"])

2. pandas straight from the Hub

If you'd rather work in tables, read the CSV directly from the Hub over the hf:// protocol (requires the huggingface_hub package):

import pandas as pd

df = pd.read_csv("hf://datasets/dmytro-yemelianov/verba/corpus.csv")

# top 10 themes
df["category"].str.split(";").explode().value_counts().head(10)

3. The hf CLI — a local copy

To download the dataset to disk (for offline processing or model training), use the official hf tool:

# install the CLI (once)
curl -LsSf https://hf.co/cli/install.sh | bash -s

# the whole dataset repo into ./verba
hf download dmytro-yemelianov/verba --repo-type dataset --local-dir ./verba

# or just a single file
hf download dmytro-yemelianov/verba corpus.csv --repo-type dataset --local-dir ./verba

4. SQL without downloading (DuckDB)

Hugging Face auto-converts the corpus to Parquet, so you can run analytical queries without fetching the whole file. Find the Parquet URLs with:

hf datasets parquet dmytro-yemelianov/verba

Then run plain SQL through DuckDB, straight from the CLI:

hf datasets sql "SELECT category, COUNT(*) AS n
  FROM read_parquet('https://huggingface.co/api/datasets/dmytro-yemelianov/verba/parquet/default/train/0.parquet')
  GROUP BY category ORDER BY n DESC LIMIT 10"

Schema and fields

One row per proverb. Columns in corpus.csv:

License and citation

• Curation and enrichment (schema, modern spelling, themes, variant groups) — CC BY 4.0.
• Historical texts (Ilkevych 1841, Nomys 1864, Franko 1901) — public domain.
• Modern collections (Bobkova, Mlodzynskyi 2009) — under their compilers' copyright; included for research and education with attribution in the sources field.

Citation string:

Yemelianov, Dmytro (2026). verba — Ukrainian Proverbs Corpus (v1.0.2). URL: https://verbacorpus.org.

What's next

• Other formats, the REST API, and citation guidelines — in “Open Data: How to Reuse the Corpus”.
• The full API documentation with curl examples.
• Source code and releases — on GitHub.