nomeda-lab/arabic-decision-benchmark

Arabic Decision Benchmark

A public benchmark for evaluating Arabic typed-decision models. Given one state (an Arabic message, record or document) and a typed question, a model must return the best option — it is a decision benchmark, not a text-generation benchmark.

It is designed so that anyone can evaluate their own model on the same data: supply the state and the candidate options, and check the prediction against the gold answer. The benchmark does not ship model predictions or results.

Tasks (22)

area	tasks
intent & scenario	banking intent, assistant intent, assistant scenario
sentiment & sarcasm	sentiment (two sources), sarcasm
rating (ordinal)	review rating (two sources)
authenticity & safety	review authenticity, hate speech
stance	stance
workflow decisions	mailroom, amount extraction, phone extraction, entity alignment, information retrieval, email selection, context retention, silent failure, sponsor segmentation, citation control, browser/agent control

Files

file	rows	contents
arabic_decision_benchmark.jsonl	5,635	the benchmark

Schema

{
  "id":        "unique row id",
  "task":      "task id",
  "family":    "task family",
  "domain":    "domain",
  "decision_type": "choice | score | noul",
  "state":     "Arabic state (message, record or document)",
  "question":  "the typed question",
  "question_lang": "ar_msa | ar_egy | ar | en",
  "language":  "ar_msa | ar_egy | ar_gulf | ar | mixed en+ar",
  "candidates": [{"label": "…", "definition": "…"}],
  "gold":      "the correct label",
  "source":    "source dataset",
  "license":   "source license"
}

• choice — select one candidate (gold is a candidate label).
• score — an ordered scale (gold is the level index as a string).
• noul — yes/no (gold is "true" or "false").

Every candidate carries a definition; gold is always one of the candidates.

Every state is Arabic. language records the variant: ar_msa / ar_egy / ar_gulf where the source annotates a dialect, ar for Arabic text, and mixed en+ar for structured Arabic content (e.g. records with English field names). Natural code-switching (product names, numbers, hashtags) is preserved as it appears in the source.

Composition

Types: choice 2,328 · noul 2,310 · score 997.

Language — two separate fields, do not confuse them:

field	meaning	breakdown
language	the state variant	ar_msa 1,216 · ar_egy 1,345 · ar_gulf 35 · ar 867 · mixed en+ar 2,172
question_lang	the question language	ar_msa 1,574 · ar_egy 706 · ar 2,161 · en 1,194

(All states are Arabic; mixed en+ar means Arabic content with English field names. question_lang = ar is used for the translated workflow questions.)

Quality and validation

Every row passed a multi-stage validation pipeline before publication:

• Structural — valid JSON, required fields present, non-empty, distinct candidate labels, gold is one of the candidates.
• Character hygiene — no CJK, no placeholder tokens (ZXQ), no control or invisible characters.
• Document/span consistency — for extraction tasks, every span's start/end resolves exactly to its value inside the document text (text[start:end] == span.text), and the value occurs in the text.
• Semantic validation — each row's question was independently re-derived and checked to be answerable from the state with the stored gold as the correct answer.
• Contamination audit — checked against the source training data (normalised exact match + near-duplicate).

The Open-Jev workflow tasks are an Arabic translation of the English Open-Jev source; their document values (ids, dates, amounts, phone numbers) are kept verbatim so each decision remains well-defined.

Reference results

For reference, three typed-decision systems were evaluated on this benchmark under an identical protocol (accuracy / macro-F1):

slice	n	QARAR	JEV-1.13	Laya
Native-Arabic tasks	3,474	0.724 / 0.671	0.696 / 0.591	0.454 / 0.386
Workflow tasks	2,161	0.626 / 0.449	0.875 / 0.747	0.426 / 0.230
Whole benchmark	5,635	0.687 / 0.560	0.765 / 0.669	0.443 / 0.308

These are provided as a sanity reference only — the benchmark is intended for anyone to evaluate their own model, and no result is required to use the data.

How to use

import json
rows = [json.loads(l) for l in open("arabic_decision_benchmark.jsonl")]

for r in rows:
    pred = my_model.predict(
        state=r["state"],
        type=r["decision_type"],
        question=r["question"],
        candidates={c["label"]: c["definition"] for c in r["candidates"]},
    )
    correct = (str(pred) == str(r["gold"]))

Report accuracy and macro-F1 overall and per task and decision_type. For score tasks, gold is the index of the correct level.

Sources and licensing

The benchmark aggregates public Arabic sources. Each row carries its source and license:

• ArBanking77 (CC-BY-SA-4.0), MASSIVE ar (CC-BY-4.0), ArSarcasm-v2 (MIT), Egyptian Fake Reviews (MIT), Egyptian Hate Speech (MIT), SHEIN Reviews (Apache-2.0), ArabicStanceX (CC-BY-4.0).
• Open-Jev (TypeSafeAI, CC0-1.0); its workflow tasks are provided here in Arabic translation.

Check each source license before redistribution. Attribution follows the original datasets.

Intended use

Evaluating Arabic typed-decision models: option selection, probability calibration and abstention. It is not a knowledge or reasoning benchmark. Do not train on it.