Pashto-OpenThoughts-15K-Reasoning

Pashto reasoning dataset based on OpenThoughts-114k, filtered to samples up to approximately 15K characters and translated into natural Pashto.

📌 Dataset Description

Pashto-OpenThoughts-15K-Reasoning is a Pashto reasoning dataset created from the OpenThoughts-114k dataset.

The dataset focuses on translating and preserving reasoning-oriented examples into Pashto while maintaining important technical structures such as:

• Python and programming code
• mathematical expressions
• LaTeX
• technical identifiers
• special reasoning markers
• problem statements
• solution explanations
• structured conversations

The goal is to provide useful Pashto reasoning and instruction-following data for training and fine-tuning Pashto-capable language models.

🧠 Source Dataset

This dataset is derived from:

OpenThoughts-114k

The source dataset contains reasoning-oriented examples covering programming, mathematics, problem solving, and other instruction-following tasks.

This project applies a character-length filter and Pashto translation while attempting to preserve code and technical content.

📊 Dataset Statistics

The original OpenThoughts dataset contained approximately 114K examples.

For this version:

• Maximum source example length: 15,000 characters
• Selected examples: 47,449
• Removed because they exceeded the 15K-character limit: 66,508
• Selected portion: approximately 41.63% of the original dataset

The dataset is organized into translation chunks to make large-scale processing and quality checking easier.

📁 Dataset Structure

Pashto-OpenThoughts-15K-Reasoning/
└── data/
    ├── chunk_001.jsonl
    ├── chunk_002.jsonl
    ├── chunk_003.jsonl
    └── ...

Each JSONL record preserves the original example structure as much as possible.

Typical fields include:

{
  "id": "000001",
  "conversations": [...],
  "system": "..."
}

🔬 Reasoning Format

OpenThoughts reasoning markers are preserved where present:

<|begin_of_thought|>
...
<|end_of_thought|>

<|begin_of_solution|>
...
<|end_of_solution|>

These markers allow downstream training pipelines to distinguish reasoning content from the final solution.

💻 Code Preservation

Programming code is treated as technical content and should remain executable.

For example:

def main():
    n = int(input())
    s = input().strip()

    red_count = s.count('R')

    print("Yes" if red_count > n - red_count else "No")

main()

Programming identifiers, syntax, operators, function names, variables, strings required by the original program, and mathematical notation should not be translated as ordinary natural language.

🧮 Mathematics and LaTeX

Mathematical expressions and LaTeX notation are preserved.

Examples:

F(x)
dp[i]
O(N)
a_i
N^3

and:

$N$
$a_i$

are treated as technical content rather than English-language text.

🌐 Language

Primary language:

Pashto (پښتو)

The dataset may naturally contain English, Japanese, mathematical notation, programming syntax, proper names, and other technical material originating from the source problems.

These elements are preserved when translation would damage their meaning or functionality.

🎯 Intended Use

This dataset can be used for:

• Pashto LLM fine-tuning
• reasoning-capable Pashto models
• instruction tuning
• supervised fine-tuning (SFT)
• code reasoning in Pashto
• mathematical reasoning in Pashto
• programming problem solving
• Pashto conversational AI
• multilingual model adaptation
• research on low-resource language reasoning

⚠️ Important Note About Reasoning Data

The reasoning content originates from a translated reasoning dataset.

Translation does not guarantee that every source example is semantically perfect.

Possible issues include:

• source-problem inconsistencies
• unusual wording
• mixed-language technical terminology
• malformed source examples
• programming statements containing English literals
• mathematical notation mixed with natural language

Therefore, users should perform additional dataset QA before high-stakes training or evaluation.

🧪 Quality Control

The translation pipeline is designed to protect:

• fenced code blocks
• inline code
• LaTeX
• URLs
• technical identifiers
• OpenThoughts reasoning markers
• programming syntax

English-leak checking should distinguish between:

Natural English language

and:

Technical/code content

For example, identifiers such as:

current_status
total_I
other_Is
append()
F(x)
dp[i]
MOD

should not automatically be classified as English-language leakage.

📦 Format

The dataset is provided in JSONL format.

Each line represents one dataset example.

This format makes the dataset suitable for:

• streaming
• incremental processing
• chunk-based training
• local preprocessing
• Hugging Face Datasets
• custom Python training pipelines

🤗 Loading the Dataset

Example:

from datasets import load_dataset

dataset = load_dataset(
    "nassimjp/Pashto-OpenThoughts-15K-Reasoning"
)

print(dataset)

To inspect examples:

print(dataset["train"][0])

🚀 Training

The dataset can be used as a source for Pashto instruction/reasoning fine-tuning.

A typical pipeline can include:

Pashto-OpenThoughts-15K-Reasoning
              ↓
        Data validation
              ↓
      Formatting / packing
              ↓
        Tokenization
              ↓
        SFT training
              ↓
      Pashto reasoning model

📝 Dataset Creation

The dataset was created by:

nassimjp

The project is focused on building useful open Pashto AI resources and increasing the availability of high-quality Pashto training data for language models.

📜 License

This repository uses the Apache-2.0 license for this dataset repository.

Users should also review the licensing and terms of the original source dataset before redistributing or using derived data commercially.

🙏 Acknowledgements

Thanks to the creators and contributors of OpenThoughts and the open-source AI community.

This dataset would not exist without the work of the original dataset authors and the broader open-source language-model ecosystem.

🔗 Repository

Hugging Face Dataset:

nassimjp/Pashto-OpenThoughts-15K-Reasoning

📌 Citation

If you use this dataset in your research or project, please cite this repository and the original OpenThoughts dataset.

@misc{nassimjp_pashto_openthoughts_15k,
  author       = {nassimjp},
  title        = {Pashto-OpenThoughts-15K-Reasoning},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/nassimjp/Pashto-OpenThoughts-15K-Reasoning}
}

---

Keywords: Pashto, پښتو, OpenThoughts, reasoning, chain-of-thought, LLM, SFT, instruction tuning, programming, mathematics, code reasoning, Pashto AI, low-resource language, synthetic dataset.