IFM/AA-LCR-1.1-AR

AA-LCR 1.1 (Arabic)

An Arabic translation of AA-LCR, the Artificial Analysis Long Context Reasoning benchmark: 100 difficult questions, each requiring reasoning over a set of real-world documents of about 100k tokens (company reports, government consultations, legal texts, academic papers and others). Answers cannot be retrieved directly and must be reasoned from several sources; responses are marked correct or incorrect by an LLM judge against the official answer. The full benchmark, questions, answers and documents, was translated into Modern Standard Arabic.

Translation

• Translated into Modern Standard Arabic with gpt-5.6-terra: questions, answers and documents.
• The 16 answer keys corrected in version 1.1 of the original are carried over; scores are therefore comparable with version 1.1, not with version 1.0.

Source

• Original dataset: ArtificialAnalysis/AA-LCR, version 1.1, revision 9a77ef56b717057ade24ceab4d273712a0b4f19e.
• Same files as the original: AA-LCR_Dataset.csv (the 100 questions) and extracted_text/AA-LCR_extracted-text.zip (the document text, under lcr/<category>/<set_id>/<filename>.txt, the same paths as the original zip). Every document referenced by a question is included (229 documents in 30 document sets); one document of the original that no question cites is omitted.

Columns

As in the original: document_category, document_set_id, question_id, question (Arabic), answer (Arabic; several accepted values are separated by ;), data_source_filenames (the documents of the set, ;-separated, in the order they are loaded into the prompt), data_source_urls and input_tokens. input_tokens is recomputed on the Arabic prompt with the cl100k_base tokenizer, following the original benchmark setup.

Prompt

The prompt template and the loading code are those of the original card: the documents of a question are concatenated in the order of data_source_filenames, each wrapped in BEGIN DOCUMENT i / END DOCUMENT i, followed by the question.

License

The source dataset is released under the Apache License 2.0, and this translation inherits it. See the original dataset card for details.

Citation

Please cite the original authors:

@dataset{artificialanalysis2025lcr,
  title={Artificial Analysis Long Context Reasoning Benchmark(LCR)},
  author={Artificial Analysis Team},
  year={2025},
  publisher={Artificial Analysis, Inc.}
}

@misc{jais3_2026,
  title  = {Jais 3: Introducing Arabic-centric Agentic AI},
  author = {{IFM Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/jais3/},
}