Arabic Multi-turn Conversational Synthetic Dataset

A synthetic dataset of 43,316 conversations with mean conversation length of 14.038 turns (rounded to 3 decimal places), median of 12 turns, range of 5-111 turns, and a total of 608,052 utterances (where every turn is an utterance).

Dataset is partitioned into training and test sets. An 80/20 split was adopted (34,653 training conversations / 8,663 test conversations).

The synthetic data generation process systematically iterated over 93 topics and 151 countries, creating 14,043 unique topic-country combinations. The generation pipeline was configured to produce 5 conversations per combination. After rigorous processing and train/test split based on techniques to mitigate leakge risks, the end result was 43,316 conversations.