APARSIN/persian-dialects-pretraining

APARSIN Low-Resource Persian Dialects Pretraining Corpus

Dataset Description

This dataset is a comprehensive pretraining corpus specifically curated for extremely low-resource Persian dialects and related regional languages. Created by the APARSIN (SilkRoadAparsin) organization, it aggregates, normalizes, and categorizes text from various sparse sources to facilitate the pretraining of Large Language Models (LLMs) on underrepresented languages and dialects.

Languages and dialects covered include (but are not limited to):

• Hazaragi
• Dari
• Gilaki
• Pashto
• Kurdish (Central/Sorani, Northern/Kurmanji, Southern)
• Standard Persian (for alignment and comparative purposes)

Dataset Structure

Data Fields

Each record in the dataset is structured to prevent schema conflicts while retaining all source metadata. The fields are:

• text (string): The raw textual content.
• source (string): The origin or source dataset of the text (e.g., Hezarai, Arman, Parsinlu, DOLMA, PEYMA, BBC).
• language (string): The specific language or dialect code (e.g., fa, glk, prs, ckb).
• extra (string): A JSON-serialized string containing additional source-specific metadata (like split, original_file, etc.).

Dataset Statistics

Language Distribution

Tokens Per Language

Usage

You can load this dataset using the Hugging Face datasets library. Because the extra field is serialized to ensure schema compatibility, you can parse it back into a dictionary using the json module.

from datasets import load_dataset
import json

# Replace with the exact repository ID where you uploaded the dataset
repo_id = "APARSIN/persian-dialects-pretraining"
dataset = load_dataset(repo_id, split="train")

# Example of reading the serialized metadata
sample = dataset[0]
extra_metadata = json.loads(sample["extra"])

print(f"Dialect: {sample['language']}")
print(f"Source: {sample['source']}")
print(f"Metadata: {extra_metadata}")
print(f"Text: {sample['text'][:100]}...")