Bayanat Labs Arabic Public Data Reasoning
Arabic Public Data Reasoning tests controlled Arabic computations over a frozen World Development Indicators snapshot: 4,400 observations across selected indicator series and 22 selected Arab League economies. Bayanat Labs generated 6,419 question rows and 4,141 visible table contexts, with exact answer programs, source-cell references and controlled English counterparts. Underlying observations and native Arabic source labels are real; the questions, annotations and table views are generated. Economy groups stay disjoint across train, validation and test. Contexts contain only each program’s evidence cells, so citation scores do not test retrieval or distractor rejection. A template-aware symbolic parser solves all released tasks; no learned-model difficulty or training benefit is established.
| Licence | CC BY 4.0 data/annotations with mandatory World Bank source terms and provider attribution; Apache-2.0 code |
|---|---|
| Version | v0.1.0 · Frozen data v0.1.0, hosted as arabic-table-reasoning. Reproduction source v0.1.1 updates destination links only; source observations, tasks, answers and splits are unchanged. |
| Published | |
| Updated | |
| Who it is for | Check Arabic numerical QA, unit handling, source-cell references and missing-value abstention against executable answers. |
| Level | Intermediate |
Key numbers
Counted by script from the released files. The SHA-256 identifies the reviewed canonical file manifest for the frozen dataset payload. Public SHA256SUMS.txt lists file checksums. Key numbers are recomputed from the exact copied summary JSON, with its own SHA-256.
Downloads and identifiers
| Where | What |
|---|---|
| Hugging Face | Versioned dataset on Hugging Face |
| Hugging Face | Explore actual released examples |
| GitHub | Versioned source code and scorer |
| huggingface.co | Frozen release statistics |
| huggingface.co | Public file checksums |
| huggingface.co | Rights, attribution and source terms |
Get started
Load the frozen dataset
from datasets import load_dataset
import json
questions = load_dataset(
'Bayanat-Labs/arabic-table-reasoning', 'questions',
split='test', revision='v0.1.0',
)
row = questions[0]
model_visible = {
'question_ar': row['question_ar'],
'table': json.loads(row['context']),
}
# Keep gold, programs, operation and diagnostic metadata out of model input.What is released
| Part | Licence | In this release |
|---|---|---|
| Selected WDI observations and source metadata | CC BY 4.0 | Yes |
| Original controlled questions, computed annotations and documentation | CC BY 4.0 | Yes |
| Original compilation and evaluation code | Apache-2.0 | Yes |
Attribution: Bayanat Labs, Arabic Public Data Reasoning, v0.1.0. Underlying data: World Bank, World Development Indicators, and the providers identified for each series. Retain provider acknowledgment and source terms; no endorsement is implied. https://www.bayanatlabs.com/open/arabic-public-data-reasoning
Intended uses
- Evaluate the declared Arabic lookup, comparison, ranking, difference and join operations using the exact displayed values.
- Test percentage-point versus relative-change handling, incompatible units and missing-value behavior.
- Reproduce source-linked answers, splits and transparent symbolic baselines from frozen bytes.
Out of scope
- Claiming that generated question rows are independent real-world facts or human-validated economic annotations.
- Using evidence scores as measurements of retrieval quality, distractor rejection or faithful internal reasoning.
- Treating the frozen observations as current forecasts or sufficient support for consequential economic decisions.
Rights at a glance
Training, display, redistribution and synthetic-voice use are separate permissions.
| Use | Status | Note |
|---|---|---|
| Training | Allowed | Permitted under CC BY 4.0 while retaining attribution, change notices, mandatory World Bank additions and provider acknowledgment; pass applicable source requirements to downstream recipients. |
| Display | Allowed | Permitted under CC BY 4.0 while retaining attribution, change notices, mandatory World Bank additions and provider acknowledgment; pass applicable source requirements to downstream recipients. |
| Redistribution | Allowed | Permitted under CC BY 4.0 while retaining attribution, change notices, mandatory World Bank additions and provider acknowledgment; pass applicable source requirements to downstream recipients. |
| Synthetic voice | Not applicable | No audio, performer recordings or voice permissions are included. |
Limitations
- Questions and annotations are software-generated controlled templates; no native-speaker or statistician validation is established. All gold and answer programs are public.
- Source facts may be estimated, revised or defined differently across economies. The selected cohort does not reconstruct historical membership or boundaries.
- Missing-value targets are rare; report answer-status results by class. Each context contains only reference evidence cells, so it does not test distractor rejection.
- Validation and test each hold out only a small economy cohort. Grammar, indicators and numeric values can recur; second-distinct-highest-year questions form a separate test-only structure diagnostic.
- The template-aware parser solves all 6,419 released tasks. Returning the cell at the start of the table performs poorly on exact answers despite high status accuracy. These are input-only rules baselines, with no LLM score or demonstrated training benefit.
Sources and licences
| Source | Licence | Redistributed here |
|---|---|---|
| World Bank — World Development Indicators and per-series providers | Eight individually licensed CC BY 4.0 series with mandatory World Bank source additions and preserved provider acknowledgment. See the exact source register; no permission for unrelated series or provider prose is implied. | Yes |
| Bayanat Labs — controlled task construction | Original generated questions, answer annotations, table assembly and documentation under CC BY 4.0; original code under Apache-2.0. | Yes |
Source observations and native Arabic economy/indicator labels come from World Bank API source 2; all eight selected series explicitly identify CC BY 4.0 in frozen metadata. API update fields are 2026-10-08; collection occurred on 2026-10-26. Bayanat Labs preserved raw responses, source hashes, provider statements, missingness and units, then added controlled questions and software answers. Keep World Bank/provider acknowledgment and the mandatory source additions, including dispute-resolution terms. No provider endorses the adaptation. English documentation is the default; Arabic source labels are preserved and this Arabic company page is optional. Source details
How to cite
Cite as:
Bayanat Labs (2026). Arabic Public Data Reasoning, v0.1.0. Bayanat Labs. https://www.bayanatlabs.com/open/arabic-public-data-reasoning
No DOI. Cite the URL in the citation above and the date you read it.
@misc{bayanatlabs2026arabicpublicdatareasoning,
author = {{Bayanat Labs}},
title = {Arabic Public Data Reasoning, v0.1.0},
year = {2026},
month = oct,
howpublished = {\url{https://www.bayanatlabs.com/open/arabic-public-data-reasoning}}
}
Changelog
| Date | Version | Change |
|---|---|---|
| v0.1.0 | Initial frozen release: 4,400 WDI observations, 6,419 controlled question rows and 4,141 visible table contexts across 22 selected economies, with exact executable answers and retained source terms. |
Corrections and removal requests
If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work
Related product: Bayanat Bench