Arabic Long-Context Conversational Memory Dataset

A synthetic dataset of 500 long multi-turn Arabic conversations with a mean conversation length of 33.056 turns, median of 33 turns, range of 24–48 turns, and a total of 16,528 turns (where every turn is one user message together with its expected agent response). Conversation length in tokens has a mean of 9,901, a median of 11,060, and a range of 710–18,456. Token counts cover the user side of each conversation and are computed under the cl100k_base tokenizer, since the dataset specifies an expected response for each turn rather than a fixed agent reply. The dataset is not partitioned. It is intended for evaluation rather than training, so all 500 conversations form a single test set. Each conversation follows one of five composition templates, allocated in equal proportion at 100 conversations each, combining personal information about a fictional persona with excerpts from Arabic Wikipedia articles and short general-knowledge exchanges. Every turn is annotated with the memory-dependent design objectives it probes, an English description of its role in the conversation, and the expected response. Across the dataset this yields 22,214 objective instances over ten design objectives, of which 16,310 query information established earlier in the same conversation. The number of turns separating a queried fact from its introduction has a median of 15 and a maximum of 47. The generation process drew on 1,200 Arabic Wikipedia articles spanning thirteen categories and totalling 4,649,321 words, with each article used in exactly one conversation, and on a curated pool of 1,244 Arabic general-knowledge topics across seventeen categories. Conversations were produced with Gemma 3 (12B) under a fixed random seed, from a plan computed deterministically in advance. Article content is derived from Arabic Wikipedia (https://ar.wikipedia.org), licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/). Articles were modified by extracting contiguous runs of whole sections of roughly 2,000 to 6,000 characters. In accordance with the ShareAlike term, this dataset is released under CC BY-SA 4.0.