rasmbench/rasmbench_revised

RasmBench

1,002 rendered Arabic documents with charts and tables, each with six questions (one per Bloom level) in multiple-choice and open-ended form: 6,012 items per format. Documents are built from Arabic Wikipedia tables in three value regimes (334 each) and five degradation tiers (A born-digital, B print scan, C complex-layout scan, D camera capture, E historical print).

• factual: the real values.
• counterfactual: every number perturbed, with totals, percentages and year order restored. Labels stay real.
• neutral: placeholder labels and random values.

Counterfactual items contain deliberately false statistics about real entities (values_altered = true). Do not use them as factual data or train on them as real data.

Configurations

Config	Splits (rows)	Contents
documents (default)	test (1,002)	The benchmark: page image, undegraded image, six questions
paired_rerenders	factual_values, counterfactual_values (668 each)	Each factual/counterfactual document drawn with both value sets, questions re-keyed
figure_as_text	image_table (100), image_table_control (50), table_only (100), own_table (230)	Figure data as a markdown table: with the image, without it, or as the model's own transcription
translation	ar_doc_ar_q, ar_doc_en_q, en_doc_ar_q, en_doc_en_q (98 each)	Arabic/English document × Arabic/English question
chart_to_table	selection (46), pairs (97), tiers (245)	Transcribe the figure; the reference is what the chart draws
model_outputs	test (201,593)	Every model answer in every condition, with judge scores
second_judge	test (4,252)	mistral-medium-3 as a second judge on 598 sampled questions

Diagnostic conditions use 100 selected documents (the 50 easiest and 50 hardest across ten models) and 50 random control documents, listed in metadata/ and in the selection column. Full page annotations are in annotations/ (re-rendered variants also carry theirs in an annotation column).

from datasets import load_dataset
docs = load_dataset("rasmbench/rasmbench_revised", split="test")
outputs = load_dataset("rasmbench/rasmbench_revised", "model_outputs", split="test")

Notes

• Key on uid (<regime>/<doc_id>), not doc_id: 21 doc_ids occur in more than one regime. In variant configs uid starts with the condition name, and source_uid points to the released document.
• Drop distractor_pool before prompting; it leaks the answer space.
• The open-ended reference answer is options[answer_index].
• tier is the degradation applied. template_nominal_tier is the template's design tier and differs on 168 documents.
• model_outputs: condition is original or the variant name, and document_uid joins to the variant's uid. Unparsed MCQ answers have a null predicted_index and count as wrong. Original open-ended answers were judged by gemini-3.5-flash and later ones by gemini-3.8-flash, so compare scores within one judge_model.

Prompts

MCQ, with options as A) … to D) …:

اقرأ الوثيقة في الصورة وأجب عن السؤال التالي.

{question}

{options}

أجب بحرف الخيار الصحيح فقط (A أو B أو C أو D).

Open-ended:

اقرأ الوثيقة في الصورة وأجب عن السؤال التالي بإيجاز.

{question}

Greedy decoding with at most 128 new tokens for MCQ, 256 for open-ended answers and 1,024 for chart-to-table transcription. The judge rubric and chart-to-table prompts are in the paper's appendix.

Limitations

• Questions were written by gemini-2.5-flash-lite and open-ended answers judged by Gemini, which may favour Gemini-like models.
• Re-rendered questions were re-keyed automatically; between 1 and 3 of 50 hand-checked were wrong.
• English documents are machine translations. 2 of the 100 selected failed to translate (metadata/translation_failed.json), and Qwen3.6-27B, gemma-4-31B, Qwen3-VL-4B and Qwen3.5-9B answered 76 of the 98.
• Tiers were not assigned at random, so tier comparisons are observational.
• The rendering engine is not part of this release.

Licence

CC BY-SA 4.0, derived from Arabic Wikipedia (CC BY-SA 4.0). Credit Arabic Wikipedia and this dataset, and keep the licence. Questions, translations and judge scores were generated with Gemini models and answers with the listed models, so their providers' terms may also apply.

Citation

@misc{rasmbench,
  title  = {RasmBench: Reading, Not Reasoning, Limits Vision-Language Models on Arabic Charts and Tables},
  author = {Anonymous},
  year   = {2026},
  note   = {Under review}
}