dataset_info:
features:
- name: text
dtype: string
- name: id
dtype: string
- name: dump
dtype: string
- name: url
dtype: string
- name: date
dtype: string
- name: file_path
dtype: string
- name: offset
dtype: int64
- name: token_count
dtype: int64
- name: language
dtype: string
- name: page_average_lid
dtype: string
- name: page_average_lid_score
dtype: float64
- name: full_doc_lid
dtype: string
- name: full_doc_lid_score
dtype: float64
- name: per_page_languages
list: string
- name: is_truncated
dtype: bool
- name: extractor
dtype: string
- name: page_ends
list: int64
splits:
- name: train
num_bytes: 12988288725
num_examples: 118768
- name: test
num_bytes: 110433769
num_examples: 1010
download_size: 6214257908
dataset_size: 13098722494
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
license: apache-2.0
task_categories:
- text-classification
- text-generation
language:
- ur
tags:
- URDU
- ur
- text
urdu_finepdfs
https://huggingface.co/datasets/HuggingFaceFW/finepdfs
This repository contains the Urdu subset of Hugging Face's FinePDFs collection — a large, multilingual corpus of parsed PDF documents. The goal of this repo is to make it easier to find, inspect, and use the Urdu portion of FinePDFs for research and model development.
NOTE: FinePDFs is maintained by HuggingFace (HuggingFaceFW). This repository is a small, focused collection and helper for the Urdu data only; it does not re-host the full FinePDFs dataset. Users should consult the original FinePDFs dataset for the full dataset card, license, and details.
What’s inside
data/(optional) — small example files / scripts [FUTURE] . This repo is primarily a pointer + helpers to the official FinePDFs Urdu shards.scripts/[FUTURE] — utility scripts to list, preview, and filter Urdu parquet shards (example: extract metadata, sample text, convert to plain text).README.md— this file.
If you cloned this repo to help with the downstream work, expect the real Urdu shards to be loaded from the official Hugging Face hub (see examples below).
About FinePDFs
FinePDFs is a massive corpus built from PDF documents. It contains hundreds of millions of PDF-derived documents across many languages and scripts and is made available by Hugging Face. The canonical dataset card, data files, and licensing are hosted on the Hugging Face dataset page for HuggingFaceFW/finepdfs.
Key facts (from the original FinePDFs project):
- FinePDFs is a multilingual PDF corpus released by Hugging Face (FinePDFs dataset). citeturn1view0turn2view0
- The dataset is very large (multiple terabytes; FinePDFs core release metadata and changelog available on Hugging Face). If you plan to work with large portions, prefer streaming or cloud compute. citeturn0search2turn1view0
- The dataset is published under the ODC-BY-style license shown on the dataset page — check the dataset card before redistributing. citeturn1view0
How to load the Urdu subset (recommended)
Option A — load the Urdu config directly (recommended if the dataset exposes a per-language config):
from datasets import load_dataset
# try loading the Urdu config by its language+script code (config name often like "urd_Arab")
dset = load_dataset("HuggingFaceFW/finepdfs", "urd_Arab", streaming=True)
# streaming=True avoids downloading large files locally; iterate and process each example
for example in dset:
print(example.get("text")[:400])
break
Option B — load the full dataset and filter (use only for small experiments or when configs are unavailable):
from datasets import load_dataset
dset = load_dataset("HuggingFaceFW/finepdfs", streaming=True)
# filter by metadata column that indicates language (often called 'language' or 'lang')
urdu = dset.filter(lambda ex: ex.get("language") in ("urd", "urd_Arab"))
for ex in urdu.take(5):
print(ex.get("text")[:200])
Notes:
- FinePDFs uses parquet shards and many per-language shard names (e.g.
urd_Arab-00000-of-00010.parquet). If a per-language config exists you can pass that config name toload_dataset. The data directory and dataset card on Hugging Face list available language/shard names. citeturn4view0turn2view0 - Prefer
streaming=Trueon large datasets to avoid downloading multi-GB/TB data.
Suggested scripts (examples in scripts/) [FUTURE]
list_shards.py— list available Urdu shards on the hub (by reading the datasetdata/directory). Useful to see exact shard filenames (e.g.urd_Arab-*.parquet).sample_text.py— stream N examples and write plain-text samples for quick inspection.parquet_to_txt.py— convert a local Urdu parquet shard to newline-separated UTF-8 text.
(Include any scripts you add here and document their CLI flags.)
License & citation
This repo is a focused helper for the Hugging Face FinePDFs Urdu subset. The original FinePDFs dataset and its license (ODC-BY, dataset card) remain the canonical source — check the dataset page for exact terms before reusing or redistributing. If you use the data in research, please cite the FinePDFs dataset and related paper(s) as described on the Hugging Face dataset card. citeturn1view0turn2view0
Contributing
If you'd like to help:
- Open an issue to request helpers or report missing shard names.
- Submit PRs that add useful scripts (e.g., better samplers, cleaning utilities) and unit tests.
- If you create smaller curated Urdu splits (cleaned, deduplicated), include exact provenance and obey the original license.
Caution & ethics
- PDFs can contain copyrighted text, PII, or sensitive information. Before using or releasing derived subsets, verify that your use case complies with the license and with ethical/data-protection requirements.
- FinePDFs is built from heterogeneous PDF sources — quality and OCR correctness vary across shards and languages.
Contact
If this repo is maintained by me (Humair) or you want to add contact info, add a MAINTAINER entry or open issues on the repo.
This README was prepared as a focused guide for the Urdu subset of Hugging Face FinePDFs.