tasal9/ZamAi-Pashto-Datasets-V2

🌐 ZamAI – Cleaned Pashto Language Dataset (V2)

پاک شوی پښتو ډیټاسیټ – د ژبني ماډلونو لپاره (نسخه ۲)

A curated, culturally aligned, and professionally cleaned Pashto dataset developed by ZamAI to accelerate open‑source NLP research for Afghanistan and the broader Pashto‑speaking world.

This release provides multiple formats (CSV + JSONL schemas), each exposed as a separate configuration to ensure schema consistency, smooth loading, and full compatibility with the Hugging Face Dataset Viewer.

---

🟦 ZamAI Identity & Branding Notes This dataset follows ZamAI’s core principles:

• Clarity — clean schemas, predictable structure
• Accessibility — open‑source, community‑driven
• Cultural fidelity — Pashto‑first, Afghan‑centric
• Technical rigor — consistent formatting, validated text, Unicode‑safe

---

🧩 Dataset Configurations

Each configuration corresponds to a single, consistent schema, as required by Hugging Face:

Config Name	Format	Schema	Ideal Use Case
default	JSONL	instruction, input, output	Instruction tuning, SFT, chat-style training
csv	CSV	text	Language modeling, embeddings, general corpus
alt_jsonl	JSONL	Alternate schema	Custom research workflows, experimentation

Why multiple configs? Because the dataset includes different file formats with different column structures, and Hugging Face requires one schema per configuration.
This ensures:

• No schema conflicts
• Smooth loading via load_dataset()
• Full Dataset Viewer compatibility

---

📊 Dataset Summary

• Language: Pashto (ps)
• Total Samples: ~28.7k cleaned entries
• Split: train (25.8k rows)
• License: CC‑BY‑4.0
• Source: Curated, cleaned, and validated by ZamAI
• Domains: General text, instructions, conversational data

---

⚡ Quick Start

Load the default instruction‑tuning JSONL `python from datasets import load_dataset

ds = load_dataset("tasal9/ZamAi-Pashto-Datasets-V2", "default") print(ds) print(ds["train"].column_names) `

Load the CSV corpus python ds = load_dataset("tasal9/ZamAi-Pashto-Datasets-V2", "csv")

Load the alternate JSONL schema python ds = loaddataset("tasal9/ZamAi-Pashto-Datasets-V2", "altjsonl")

---

🧼 Cleaning & Processing Pipeline

All samples were processed using ZamAI’s internal text‑cleaning workflow:

• Unicode normalization
• Removal of HTML artifacts
• Deduplication
• Correction of broken Pashto glyphs
• RTL rendering validation
• Filtering of extremely short or noisy entries

This ensures high‑quality, model‑ready Pashto text.

---

🧠 Intended Use

This dataset is ideal for:

• Pashto LLM training
• Instruction tuning (SFT)
• Text generation
• Classification & sentiment analysis
• Question answering
• Summarization
• Cultural & linguistic preservation research

---

🤝 Citation

@dataset{zamaipashtov2, title = {ZamAI – Cleaned Pashto Language Dataset (V2)}, author = {ZamAI}, year = {2025}, url = {https://huggingface.co/datasets/tasal9/ZamAi-Pashto-Datasets-V2} }

---

🟦 About ZamAI

ZamAI builds open‑source AI tools, datasets, and research resources to support:

• Afghan languages
• Digital preservation
• Community‑driven NLP innovation
• Accessible AI for Afghanistan

Our mission is to empower the next generation of Afghan technologists through open, high‑quality AI infrastructure.

---