nassimjp/pashto-fallacy-dataset

Pashto Fallacy Dataset (د پښتو منطقي تېروتنو ډاټاسیټ)

The Pashto Fallacy Dataset is a high-quality, linguistically curated corpus containing 2,154 atomic instruction-tuning pairs. It is engineered specifically to train large language models (LLMs) to detect, classify, and logically refute informal reasoning fallacies within Pashto-centric contexts.

The dataset utilizes the standard Alpaca format (instruction, input, output), making it plug-and-play compatible with fine-tuning platforms like LLaMA-Factory, Unsloth, Axolotl, and TRL.

---

Dataset Layout & Schema

Each entry in the JSON array contains three core structural fields:

• instruction (System Prompt): A localized, explicit command in Pashto instructing the model to act as a logical analyst.
• input (Contextual Prompt): Combines a dynamic topic tag (موضوع:) and the targeted argument text containing the logical flaw (متن:).
• output (Gold Target Response): A clean, structurally isolated logical refutation (وضاحت) highlighting why the premises fail to support the conclusion.

Schema Example

{
  "instruction": "په لاندې ورکړل شوي متن کې منطقي تېروتنه (Fallacy) وپېژنئ او په منطقي توګه وضاحت ورکړئ چې ولې دا استدلال ناسم دی.",
  "input": "موضوع: د ښځو کار\nمتن: ما دوه ښځې ولیدلې چې کار کوي او بریالۍ دي، نو ټولې ښځې باید کار وکړي.",
  "output": "له لږو بېلګو څخه عمومي پایله نشي کېدای."
}

---

Technical Specifications & Curation Standards

• Multi-Domain Scope: Covers over 100 distinct contemporary, regional, and global conversational domains (e.g., society, climate change, artificial intelligence, healthcare, and education).
• Zero Target Bleeding: Unlike naive stream-parsing scripts which misalign lines when topic headers shift, this dataset was compiled using explicit atomic block segmentation. Answers and contexts are 100% paired without data leaks.
• Fallback Routing: Instances lacking an explicit sub-topic declaration drop back cleanly to a standardized general domain tag (موضوع: عمومي), ensuring zero token pollution or empty contexts during tokenization.
• Token Sanitization: Stripped completely of rogue markdown syntax anomalies (``), bracket leakage, and hanging tailing spaces to ensure smooth compute passes.

Intended Use Cases

1. Supervised Fine-Tuning (SFT): Enhancing the critical thinking, reasoning steps, and alignment behavior of open-source conversational models in low-resource language tracks.
2. Evaluation Benchmarking: Serving as an analytical check to measure the truthfulness and logical reasoning capabilities of Pashto-capable language models.

---

Licensing & Distribution

This dataset is licensed under the Apache License 2.0.

• Commercial Use: Permitted. You can use this dataset to train commercial models, host corporate services, or bundle downstream software without royalty obligations.
• Modification & Distribution: Permitted, provided proper attribution is maintained and license states are preserved.
• Liability: Distributed on an "AS IS" basis, without warranties or conditions of any kind.