HeshamHaroon/asfar-grpo

Asfar-GRPO是一个规则可验证的GRPO训练数据集，专为古典阿拉伯语设计。数据集包含76,085行数据，涵盖6种任务类型，每种任务的答案都经过机械验证以确保与源页面的准确性。数据集来源于公共领域的阿拉伯遗产文本（包括语法、词典学、文学和半岛历史），并支持多种NLP任务，如文本生成、问答、填空和多项选择。数据集采用CC-BY-4.0许可，所有样本均经过严格验证，确保质量。

Asfar-GRPO is a rule-verifiable GRPO training dataset for classical Arabic. It consists of 76,085 rows across 6 task types, with each samples answer mechanically verified against the source page. The dataset is derived from a corpus of public-domain Arabic heritage texts (grammar, lexicography, adab, peninsula history) and supports various NLP tasks such as text-generation, question-answering, fill-mask, and multiple-choice. Licensed under CC-BY-4.0, the dataset ensures high-quality samples through rigorous verification processes.