lgi2p/finefreq

FineFreq is a large-scale multilingual character frequency dataset derived from 96.6 trillion characters across 1900+ languages, built from the FineWeb and FineWeb2 corpora. It provides per-language frequency tables with year-level temporal resolution, covering the period 2013–2025.

---

Dataset Structure

Each file (per language-script pair) contains fields as follows:

Column	Description
iso_639_3	ISO 639‑3 language code
script	Unicode script (e.g. Latn, Cyrl, Hani)
language_code	Combined language-script tag (e.g. eng_Latn)
language_name	Human-readable language name
character	Unicode character
unicode_category	Unicode general category (e.g. Ll, Lo, Pd)
unicode_name	Official Unicode character name
total_frequency_all_time	Total count of the character across all available years
time_periods_with_data_count	Number of years with observed data
time_periods_list	List of years in which the character appeared
year_20XX_frequency	Character frequency per year (e.g. year_2017_frequency)

Not all languages have full year coverage. Frequencies are raw counts (not percentages), normalized later as needed.

Two manifest files are included for quick indexing:

• manifest/languages.csv
• manifest/languages.parquet

Each row contains metadata per language-script pair:

Field	Description
iso_639_3	ISO 639‑3 code
language_code	Language-script ID
script	Script used
language_name	Full language name
total_frequency	Sum of all character frequencies
data_source	FineWeb or FineWeb2
years_count	Number of years with data
min_year, max_year	First and last year of data coverage

Usage

Load a single language dataset (e.g. English)

Use the standard Hugging Face load_dataset function to load data for any language.

from datasets import load_dataset

# Load character frequency data for English (eng_Latn)
ds = load_dataset(
    "lgi2p/finefreq",
    data_files="DATA/eng_Latn/eng_Latn.parquet",  # Use .csv for CSV format
    split="train"
)

# Convert to pandas DataFrame for analysis
df = ds.to_pandas()

# Display the top 20 characters by total frequency
df_sorted = df.sort_values("total_frequency_all_time", ascending=False)
print(df_sorted[["character", "unicode_name", "total_frequency_all_time"]].head(20))

---

Browse available languages

Before loading data, you can check which languages are available in the dataset.

import pandas as pd
from huggingface_hub import hf_hub_download

# Download and read the manifest file
manifest_path = hf_hub_download(
    repo_id="lgi2p/finefreq",
    filename="manifest/languages.csv",
    repo_type="dataset"
)
manifest_df = pd.read_csv(manifest_path)

# Show a summary of available languages
print(f"Total languages available: {len(manifest_df)}")
print("\nLanguages with the most character occurrences:")
print(manifest_df.nlargest(10, 'total_frequency')[['language_code', 'script', 'total_frequency']])

---

Optional: Helper function to load any language by code

def load_finefreq(language_code: str, format="parquet"):
    """
    Load a FineFreq language dataset.
    
    Args:
        language_code: Language identifier (e.g., 'eng_Latn', 'fra_Latn')
        format: File format ('parquet' or 'csv')
    
    Returns:
        pandas.DataFrame with character frequency data
    """
    from datasets import load_dataset
    ext = "csv" if format == "csv" else "parquet"
    ds = load_dataset(
        "lgi2p/finefreq",
        data_files=f"DATA/{language_code}/{language_code}.{ext}",
        split="train"
    )
    return ds.to_pandas()

# Example usage
df_english = load_finefreq("eng_Latn")
df_french = load_finefreq("fra_Latn")

---

Notes

• Replace "eng_Latn" with any language code listed in manifest/languages.csv.
• Both .parquet and .csv files are available for each language.
• Parquet is recommended for faster loading; CSV is human-readable and previewable on the Hugging Face Hub.
• Each language folder also contains a metadata.json file with summary statistics and source information.

---

Citation

@article{Xu2025finefreq,
  title={FineFreq: A Multilingual Character Frequency Dataset from Web-Scale Text},
  author={Binbin Xu},
  journal={arXiv preprint arXiv:2512.09701},
  year={2025}, 
  url = {https://arxiv.org/abs/2512.09701}
}