nassimjp/pashto-clean-conversations

Pashto Clean Conversations Dataset 🇦🇫✨

This is a highly curated, deduplicated, and linguistically sanitized conversational dataset in Pashto. It is specifically structured for instruction tuning, alignment, and perfecting the conversational capabilities of local Large Language Models (LLMs) in the Pashto language.

Dataset Summary

• Total Conversations: 722 unique multi-turn/single-turn interactions (derived from 907 raw instances).
• Format: OpenAI / Hugging Face standard messages format (role and content).
• Language: 100% Pashto (پښتو) text.
• Data Quality: 185 duplicates removed (~20% noise reduction). Pure machine-translated artifacts and sub-standard spelling errors (like پښتونانو, پاشتو, پشتو) have been algorithmically hunted and corrected into pure, native standard Pashto.

Structure & Schema

Each row is formatted as a JSON object inside a .jsonl file with a top-level messages array, making it natively compatible with Unsloth, Hugging Face Trainer, and Axolotl.

Example Entry:

{
  "messages": [
    {"role": "user", "content": "سلام علیکم ماښام مو پخیر"},
    {"role": "assistant", "content": "وعلیکم السلام، ماښام مو پخیر! زه څنګه کولای شم نن ستاسو کومک وکړم؟"}
  ]
}

Data Cleansing & Sanitization Process

1. Aggregation: Merged 10 distinct conversational JSON source tracks into a single unified workspace.
2. Deduplication: Hashed exact string matches of the conversation bodies to eliminate duplicate training patterns.
3. Linguistic Deep Hunt: Executed a specialized Regex pipeline targeting the tokenization anomalies and keyboard errors often generated by foreign local models when rendering Pashto script.

• Non-standard plurals (e.g., پښتونانو corrected to پښتنو).
• Persian/Urdu script overlaps (e.g., پشتو corrected to پښتو).
• Phonetic translation spelling errors (e.g., پاشتو corrected to پښتو).

Usage & Unsloth Integration

You can easily load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("nassimjp/pashto-clean-conversations", split="train")

For Unsloth users applying a chat template (e.g., Llama-3 or ChatML):

from unsloth import get_chat_template

dataset = get_chat_template(
    dataset,
    chat_template = "llama-3",
    mapping = {"role" : "role", "content" : "content", "user" : "user", "assistant" : "assistant"},
    mapping_column = "messages",
)

License

This dataset is distributed under the Apache-2.0 license. Feel free to use it for training and refining your Pashto AI architectures!