nassimjp/perfect-pashto-reasoning-sft

该数据集是基于Magpie-Pro-300K-Filtered数据集构建的，专为普什图语（Pashto）大语言模型和AI社区全新设计和处理的高质量推理数据集。所有数据以原子和逐行方式翻译，并以高质量重新格式化，以确保与Hugging Face的alignment-handbook直接兼容。其目标是为普什图语新一代模型（如Rawan和Ghanam系列）提升监督微调（SFT）和深度推理（链式思考）能力。数据集包含300,000个样本，其中训练集270,000个，测试集30,000个，格式为JSONL。每个记录包含system、instruction和output三个关键字段，分别表示模型身份和推理框架、用户普什图语问题或任务、以及标准、流畅、高质量的普什图语逻辑答案。数据集经过去重、原子对齐、智能分块和高质量控制等处理，适用于监督微调、链式思考、普什图语指令遵循、学术、技术和创意写作以及合成数据生成等领域。

This dataset is built based on the Magpie-Pro-300K-Filtered dataset, specifically engineered and processed from scratch for the Pashto LLM and AI community. All data has been translated atomically and line-by-line, and reformatted with high quality to directly align with the alignment-handbook. Its goal is to enhance Supervised Fine-Tuning (SFT) and Deep Reasoning (Chain-of-Thought) capabilities for new-generation Pashto models (such as the Rawan and Ghanam series). The dataset contains 300,000 examples, with 270,000 for training and 30,000 for testing, in JSONL format. Each record includes three key fields: system (model identity and reasoning framework), instruction (user query or task in Pashto), and output (logical answer in standard, fluent, and high-quality Pashto). It has undergone deduplication, atomic alignment, intelligent chunking, and high-quality control, and is suitable for Supervised Fine-Tuning, Chain-of-Thought reasoning, Pashto instruction following, academic, technical, and creative writing, and synthetic data generation.