Coverage correction: wrong-language configs, unread nested text, re-sample

`results.jsonl` (the DOI record) is untouched.

- `scan_hf.py`: prefers Arabic-coded configs, reads nested strings, reports `arabic_fields`, imports the installed arabic-lint
- New `rescan_configs.py`; re-sample of 39 datasets in `results_rescan_2026-10.jsonl` and `rescan_summary.md` (run on 27 Oct with a browser port of the same rules, cross-checked against arabic-lint 0.8.0)
- **New finding:** `HPLT/DocHPLT` and `HuggingFaceFW/finepdfs-edu` carry `reshaped` presentation-form runs in a small share of sampled fields, in logical order; the card's 'only one reshaped dataset' caveat is updated
- Card: coverage correction (16 non-Arabic configs + 3 ace_Arab, 21 zero-field datasets), correction-log entry, reproduction note, re-sample limits stated
- Ships `targets.json` so `audit_all.py` runs; MIT `LICENSE` for the code; removed the `/tmp/al4` path from add_severity.py; removed `task_categories: text-classification`

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GaKH7wkxcBhgmbNVzYaxR2

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01GaKH7wkxcBhgmbNVzYaxR2

LICENSE

@@ -0,0 +1,30 @@
+MIT License
+
+Copyright (c) 2026 Syamjith NK
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+---
+
+NOTE ON SCOPE
+
+The CODE in this repository (scorers, harnesses, generators) is MIT licensed as
+above. The DATA and RESULTS (test sets, per-item results, rendered outputs) are
+released under Creative Commons Attribution 4.0 International (CC BY 4.0), as
+declared in the dataset card.

README.md

@@ -2,8 +2,6 @@
 license: cc-by-4.0
 language:
   - ar
-task_categories:
-  - text-classification
 tags:
   - arabic
   - data-quality
@@ -151,6 +149,44 @@ SentencePiece model is immune, because XLM-R normalises and they do not. Re-runn
   only; nothing in this dataset measures visual order, and no row here should be read as a
   claim about it.
 
+## Coverage correction, 27 October 2026
+
+`scan_hf.py` took the first config that had a `train` or `test` split, without
+checking its language. For a multilingual corpus that is usually an alphabetically
+early, non-Arabic config, so **16 of the 276 readable datasets were sampled in another
+language** — for example `allenai/c4` as `af`, `wikimedia/wikipedia` as
+`20231101.ab`, `gsarti/flores_101` as `afr` and `CohereLabs/Global-MMLU` as `am` —
+and 3 more as Acehnese in Arabic script (`ace_Arab`). Those 19 contribute 8,534 of
+the 119,517 text fields. Separately, **21 readable datasets contributed 0 text
+fields**, because the scanner read only top-level string columns and their text sits
+in lists or nested records. About 240 of the 276 were therefore scanned in Arabic.
+
+This also corrects *What this does not establish*: several of the largest corpora
+(C4, Wikipedia, HPLT, FLORES) were readable. They were sampled in the wrong language,
+not blocked by the viewer.
+
+`results.jsonl` is left exactly as published, because it is the record behind the
+DOI. `scan_hf.py` now prefers a config whose name carries an Arabic language code,
+reads strings inside lists and records, and reports `arabic_fields`.
+`rescan_configs.py` re-samples the affected datasets into
+`results_rescan_2026-10.jsonl`.
+
+Re-sampled on 27 October 2026 with arabic-lint 0.8.0's detection rules, 100 rows per dataset:
+
+- Re-sampled: **39** datasets (18 on the wrong config, 21 with zero fields read); 4 could not be read this time.
+- Read now: **526,366** text fields, of which **440,918** contain Arabic letters.
+- Datasets with any finding: **3** — spans by severity: {'reshaped': 6926, 'partial': 4846, 'stray': 1873}.
+  - `HPLT/DocHPLT` (ar-en): 7/13100 fields, worst severity reshaped
+  - `HuggingFaceFW/finepdfs-edu` (arb_Arab): 11/4310 fields, worst severity reshaped
+  - `allenai/c4` (ar): 3/300 fields, worst severity stray
+- Still no text fields (audio, images or non-text data): `heshamelmahdi/egyptian-arabic-orpheus-tokenized`.
+
+Of the 18 config switches, 15 move a wrong-language sample to an Arabic config (`sarulab-speech/yodas2_sidon` could not be read this time), and `Helsinki-NLP/opus-100` and `unimelb-nlp/wikiann` were re-sampled with the zero-field group. The other three switches are not language fixes: `MBZUAI/ArabicMMLU` names its configs by subject and is Arabic either way; `quranlab/quran-audio` matched the reciter name `hani-ar-rifai`, and its new sample holds no Arabic letters; `quranlab/hadith` moved from its `index` config to an Arabic one but could not be read. `QCRI/ImageEval-ArabicNLP26` (Arabic is `task1a_msa`) and `mhardalov/exams` (Arabic only inside `multilingual`) have no Arabic-coded config name, so they were not re-sampled.
+
+Nested strings now count one field each, so token lists dominate the total (`occiglot/tokenizer-wiki-bench` alone contributes 487,296); compare finding rates, not field totals, with the September run.
+
+**What the re-sample changes.** Two large PDF-derived corpora, `HPLT/DocHPLT` and `HuggingFaceFW/finepdfs-edu`, carry `reshaped` runs of stored presentation forms in a small share of sampled fields (7 of 13,100 and 11 of 4,310). The sampled runs are in logical order, which is the pattern [ArPDF](https://huggingface.co/datasets/syamjithnk/arpdf) measures when text is extracted from a PDF that stores shaped glyphs; this sample does not establish the cause. It does change the caveat above: `Yousefmd/arabic_ocr_dataset` is no longer the only dataset with `reshaped` runs, and stored presentation forms reach two widely used corpora, at a low rate.
+
 ## Corrections
 
 Findings here are about other people's public work, so there has to be a way to contest them
@@ -173,6 +209,17 @@ wrongly is enough; you do not need to diagnose the bug.
 
 ### Correction log
 
+**2026-10-27 — coverage: wrong-language configs and unread fields.** Found in an
+external audit of this card. 16 readable datasets were sampled on a non-Arabic
+config, 3 on `ace_Arab`, and 21 yielded no text fields; see *Coverage correction*.
+Cause: the split chooser assumed a dataset's first `train`/`test` config was its
+Arabic one, and the scanner skipped nested text. The headline finding, one dataset
+with presentation forms in every sampled field, does not depend on either. Fixed in
+the same pass: `targets.json` now ships, so `audit_all.py` runs as published; the
+scripts import the installed `arabic-lint` instead of a scratch path; the code
+is licensed MIT (`LICENSE`); and `task_categories: text-classification` was removed,
+because this dataset is not a classification task.
+
 **2026-09-13 — `Yousefmd/arabic_ocr_dataset` is not stored in visual order.** The card, the
 findings write-up and a discussion opened on that repository all stated that the labels had
 been through `arabic_reshaper.reshape()` **and** `bidi.algorithm.get_display()`, and called
@@ -208,6 +255,10 @@ pip install arabic-lint
 python3 audit_all.py          # https://github.com/Syamjith-NK/arabic-lint
 ```
 
+`targets.json` (the 341 dataset ids) ships alongside since 27 October 2026;
+before that `audit_all.py` could not run as published. `rescan_configs.py`
+reproduces the coverage correction.
+
 ## Citation
 
 Archived with a DOI on Zenodo, so the exact version behind these numbers stays citable even
@@ -223,3 +274,5 @@ as this card is updated: **[10.5281/zenodo.22733934](https://doi.org/10.5281/zen
   url       = {https://doi.org/10.5281/zenodo.22733934}
 }
 ```
+
+**License:** data CC BY 4.0; code MIT (see `LICENSE`, added 27 October 2026).

add_severity.py

@@ -7,7 +7,6 @@ recorded how MANY findings, never what KIND. A published number that cannot be i
 is just an assertion with a filename.
 """
 import json, sys, time, urllib.parse, urllib.request, collections, pathlib
-sys.path.insert(0, "/tmp/al4/src")
 from arabic_lint.detect import scan_text, SEVERITY_ORDER
 
 def rows(ds, cfg, sp, n=100, tries=3):

rescan_configs.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Re-sample the datasets that the September run scanned in the wrong place.
+
+Author: Syamjith NK
+
+Two groups from results.jsonl are re-sampled with the corrected scan_hf.audit():
+
+  1. wrong-config: the stored config is not "default" and not Arabic-coded, and the
+     dataset now offers an Arabic-coded config, so the sample came from another
+     language (allenai/c4 as "af", wikimedia/wikipedia as "20231101.ab", ...).
+  2. zero-fields: readable, but 0 text fields were read, because the text sits in
+     lists or nested records.
+
+results.jsonl is NOT modified: it is the record behind the DOI. Output goes to
+results_rescan_2026-10.jsonl and rescan_summary.md.
+
+    pip install arabic-lint
+    python rescan_configs.py
+"""
+import collections, json, time
+from pathlib import Path
+
+from scan_hf import audit, choose_split, is_arabic_config
+
+HERE = Path(__file__).parent
+rows = [json.loads(l) for l in (HERE / "results.jsonl").read_text(encoding="utf-8").splitlines()
+        if l.strip()]
+
+todo = []
+for r in rows:
+    if not r.get("readable"):
+        continue
+    if r["text_fields"] == 0:
+        todo.append((r, "zero-fields"))
+    elif r["config"] != "default" and not is_arabic_config(r["config"]):
+        cfg, _, rule, err = choose_split(r["dataset"])
+        time.sleep(0.5)
+        if not err and rule == "arabic-config" and cfg != r["config"]:
+            todo.append((r, "wrong-config"))
+
+print(f"re-sampling {len(todo)} datasets", flush=True)
+out = []
+for i, (old, why) in enumerate(todo, 1):
+    new = audit(old["dataset"], n=100)
+    new.update({"reason_rescanned": why, "previous_config": old["config"],
+                "previous_text_fields": old["text_fields"],
+                "previous_fields_with_findings": old["fields_with_findings"]})
+    out.append(new)
+    tag = "ERROR " + new["error"][:50] if "error" in new else (
+        f"{new['fields_corrupted']}/{new['text_fields']} fields, "
+        f"{new['arabic_fields']} Arabic, config {new['config']}")
+    print(f"  [{i}/{len(todo)}] {why:12} {old['dataset'][:50]:52} {tag}", flush=True)
+    time.sleep(1.0)
+
+(HERE / "results_rescan_2026-10.jsonl").write_text(
+    "".join(json.dumps(o, ensure_ascii=False) + "\n" for o in out), encoding="utf-8")
+
+ok = [o for o in out if "error" not in o]
+hits = [o for o in ok if o["fields_corrupted"]]
+sev = collections.Counter()
+for o in hits:
+    sev.update(o.get("severity") or {})
+groups = collections.Counter(o["reason_rescanned"] for o in out)
+
+lines = [
+    f"- Re-sampled: **{len(out)}** datasets ({groups['wrong-config']} on the wrong "
+    f"config, {groups['zero-fields']} with zero fields read); "
+    f"{len(out) - len(ok)} could not be read this time.",
+    f"- Read now: **{sum(o['text_fields'] for o in ok):,}** text fields, of which "
+    f"**{sum(o['arabic_fields'] for o in ok):,}** contain Arabic letters.",
+    f"- Datasets with any finding: **{len(hits)}**"
+    + (f" — spans by severity: {dict(sev)}" if hits else "") + ".",
+]
+for o in hits:
+    lines.append(f"  - `{o['dataset']}` ({o['config']}): {o['fields_corrupted']}/"
+                 f"{o['text_fields']} fields, worst severity "
+                 f"{max(o['severity'], key=lambda s: ['stray', 'partial', 'reshaped'].index(s)) if o.get('severity') else 'n/a'}")
+still_zero = [o["dataset"] for o in ok if o["text_fields"] == 0]
+if still_zero:
+    lines.append(f"- Still no text fields (audio, images or non-text data): "
+                 + ", ".join(f"`{d}`" for d in still_zero) + ".")
+(HERE / "rescan_summary.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
+print("\n" + "\n".join(lines))

rescan_summary.md

@@ -0,0 +1,7 @@
+- Re-sampled: **39** datasets (18 on the wrong config, 21 with zero fields read); 4 could not be read this time.
+- Read now: **526,366** text fields, of which **440,918** contain Arabic letters.
+- Datasets with any finding: **3** — spans by severity: {'reshaped': 6926, 'partial': 4846, 'stray': 1873}.
+  - `HPLT/DocHPLT` (ar-en): 7/13100 fields, worst severity reshaped
+  - `HuggingFaceFW/finepdfs-edu` (arb_Arab): 11/4310 fields, worst severity reshaped
+  - `allenai/c4` (ar): 3/300 fields, worst severity stray
+- Still no text fields (audio, images or non-text data): `heshamelmahdi/egyptian-arabic-orpheus-tokenized`.

results_rescan_2026-10.jsonl

@@ -0,0 +1,39 @@
+{"dataset": "ApexVOrteX-1/egyptian-arabic-style-dataset", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 659, "arabic_fields": 219, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "CohereLabs/Global-MMLU", "config": "ar", "split": "test", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 1600, "arabic_fields": 358, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "am", "previous_text_fields": 1600, "previous_fields_with_findings": 0}
+{"dataset": "CohereLabs/include-base-44", "config": "Arabic", "split": "test", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 1100, "arabic_fields": 495, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "Albanian", "previous_text_fields": 1000, "previous_fields_with_findings": 0}
+{"dataset": "CohereLabs/xP3x", "config": "acm_Arab", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 700, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ace_Arab", "previous_text_fields": 700, "previous_fields_with_findings": 0}
+{"dataset": "Davlan/sib200", "config": "acm_Arab", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ace_Arab", "previous_text_fields": 200, "previous_fields_with_findings": 0}
+{"dataset": "Falah/classification_arabic_dialects", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "HPLT/DocHPLT", "config": "ar-en", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 13100, "arabic_fields": 2319, "fields_corrupted": 7, "spans": 16, "severity": {"reshaped": 4, "partial": 6, "stray": 6}, "example": "ﻟﻌﻘﺪ ﺍﻟﺘﺴﻭﻴﺔ ﺍﻟﻘﻀﺎﺌﻴﺔ ﺃﻭ ﺍﻟﺘﺼﻔﻴﺔ ﺍﻟﻘﻀﺎﺌﻴﺔ ﻟﻠﻌﻀﻭ ﺍﻟﻤﻜﻠﻑ ﺒﺎﻟﻤﻘ", "pct": 0.05, "reason_rescanned": "wrong-config", "previous_config": "af-en", "previous_text_fields": 300, "previous_fields_with_findings": 0}
+{"dataset": "HPLT/HPLT2.0_cleaned", "config": "ara_Arab", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 5029, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ace_Arab", "previous_text_fields": 144, "previous_fields_with_findings": 0}
+{"dataset": "Helsinki-NLP/opus-100", "config": "ar-de", "split": "test", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "af-en", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "HuggingFaceFW/finepdfs-edu", "config": "arb_Arab", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 4310, "arabic_fields": 100, "fields_corrupted": 11, "spans": 13620, "severity": {"reshaped": 6922, "stray": 1858, "partial": 4840}, "example": "ﻜﻨﻴﺴﺔ ﺍﻟﺴﻴﺩﺓ ﺍﻟﻌﺫﺭﺍﺀ ﺒﺎﻟﻔﺠﺎﻟﺔ", "pct": 0.26, "reason_rescanned": "wrong-config", "previous_config": "eng_Latn", "previous_text_fields": 1000, "previous_fields_with_findings": 0}
+{"dataset": "HumynLabs/Arabic_Documents_Dataset_PDF", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "MBZUAI/ArabicMMLU", "config": "Arabic Language (General)", "split": "test", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 992, "arabic_fields": 592, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "Accounting (University)", "previous_text_fields": 774, "previous_fields_with_findings": 0}
+{"dataset": "Nooh/arabic-asr-audio", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "allenai/c4", "config": "ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 300, "arabic_fields": 100, "fields_corrupted": 3, "spans": 9, "severity": {"stray": 9}, "example": "ﭙ", "pct": 1.0, "reason_rescanned": "wrong-config", "previous_config": "af", "previous_text_fields": 300, "previous_fields_with_findings": 0}
+{"dataset": "bigscience/xP3all", "config": "ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 141, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ak", "previous_text_fields": 200, "previous_fields_with_findings": 0}
+{"dataset": "bigscience/xP3mt", "config": "ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 149, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ak", "previous_text_fields": 200, "previous_fields_with_findings": 0}
+{"dataset": "cloudaocr/arabic-synthetic-scanned-books", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "craneset/arabic-ocr", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 8, "text_fields": 8, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "disco-eth/WorldSpeech", "config": "ar_bh", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 1000, "arabic_fields": 200, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "ca_fr", "previous_text_fields": 800, "previous_fields_with_findings": 0}
+{"dataset": "emadjumaah/mishkat-quran-audio", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "facebook/xnli", "config": "ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 200, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "all_languages", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "gagan3012/arabic-news-clustering-p2p", "error": "TypeError: Failed to fetch", "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "gsarti/flores_101", "config": "ara", "split": "dev", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 399, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "afr", "previous_text_fields": 399, "previous_fields_with_findings": 0}
+{"dataset": "heshamelmahdi/egyptian-arabic-orpheus-tokenized", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 0, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "kokojake/lmsys_egyptian_arabic_convs", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 736, "arabic_fields": 347, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "kokojake/oasst2_egyptian_arabic_convs", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 2256, "arabic_fields": 1112, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "melsiddieg/qari-arabic-ocr-10k", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 500, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "mustaphaelkady/arabic-ocr-books", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 500, "arabic_fields": 100, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "occiglot/tokenizer-wiki-bench", "config": "ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 487296, "arabic_fields": 433050, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "af", "previous_text_fields": 400, "previous_fields_with_findings": 0}
+{"dataset": "quranlab/hadith", "error": "HTTPError: HTTP Error 500", "reason_rescanned": "wrong-config", "previous_config": "index", "previous_text_fields": 500, "previous_fields_with_findings": 0}
+{"dataset": "quranlab/quran-audio", "config": "hani-ar-rifai", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 1300, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "manifest", "previous_text_fields": 900, "previous_fields_with_findings": 0}
+{"dataset": "saleh-c4/arabic-ocr-images", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 100, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "sarulab-speech/yodas2_sidon", "error": "HTTPError: HTTP Error 500", "reason_rescanned": "wrong-config", "previous_config": "aa000", "previous_text_fields": 92, "previous_fields_with_findings": 0}
+{"dataset": "tamm5y5m5/quran_nemotron_dual_v5_npl", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "unimelb-nlp/wikiann", "config": "ar", "split": "test", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 1481, "arabic_fields": 640, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "ace", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "wikimedia/wikipedia", "config": "20231101.ar", "split": "train", "config_rule": "arabic-config", "rows_sampled": 100, "text_fields": 400, "arabic_fields": 196, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "wrong-config", "previous_config": "20231101.ab", "previous_text_fields": 399, "previous_fields_with_findings": 0}
+{"dataset": "zaibihassan/Quranic-Recitation-Data", "error": "HTTPError: HTTP Error 501", "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "zaibihassan/Quranic-Translation-Audio-Data", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}
+{"dataset": "zaibihassan/Quranic-Word-By-Word-Audio-Data", "config": "default", "split": "train", "config_rule": "single-config", "rows_sampled": 100, "text_fields": 200, "arabic_fields": 0, "fields_corrupted": 0, "spans": 0, "severity": null, "example": null, "pct": 0.0, "reason_rescanned": "zero-fields", "previous_config": "default", "previous_text_fields": 0, "previous_fields_with_findings": 0}

scan_hf.py

@@ -1,68 +1,148 @@
 #!/usr/bin/env python3
-"""Audit public Arabic datasets for text that was corrupted before it was stored.
+"""Audit public Arabic datasets for presentation forms stored in their text.
 
-The question this answers has not been asked: models are trained and evaluated on
-Arabic corpora, and the reshape+bidi recipe silently writes presentation-form,
-visually-ordered Arabic into files. Where that happened upstream of a dataset, every
-model trained or scored on it is learning from text no Arabic reader would accept.
+Author: Syamjith NK
 
 Method: pull rows through the HF datasets-server (no full download), run the same
-detector that ships in arabic-lint, and report the share of rows affected. Sampled,
-not exhaustive, and the sample size is reported with every number.
+detector that ships in arabic-lint, and report the share of text fields affected.
+Sampled, not exhaustive, and the sample size is reported with every number.
+
+    pip install arabic-lint
+    python scan_hf.py MBZUAI/ArabicMMLU
+
+Corrected 2026-10-27:
+
+- Config choice. The September run took the first config with a `train` or `test`
+  split. For a multilingual corpus that is usually an alphabetically early,
+  non-Arabic config, so `allenai/c4` was sampled as Afrikaans and
+  `wikimedia/wikipedia` as Abkhaz. `choose_split` now prefers a config whose name
+  carries an Arabic language code, and records which rule it used.
+- Nested text. Only top-level string columns were read, so datasets that keep their
+  text in lists or records (conversations, translation pairs, token lists)
+  contributed zero fields. `_strings` now walks lists and dicts.
+- Arabic content. `arabic_fields` counts the fields that contain at least one Arabic
+  letter, so a sample in the wrong language is visible in the output.
+- The detector is imported from the installed package rather than a scratch path.
+
+`first_split` is kept unchanged because it is what produced `results.jsonl`.
 """
-import json, sys, urllib.parse, urllib.request
+import json, re, sys, time, urllib.parse, urllib.request
 
-sys.path.insert(0, "/tmp/al2/src")
 from arabic_lint.detect import scan_text
 
 API = "https://datasets-server.huggingface.co"
-UA = {"User-Agent": "arabic-lint-audit/1.0"}
+UA = {"User-Agent": "arabic-lint-audit/1.1"}
+
+# ISO 639 codes for Arabic and its varieties as they appear in config names
+# ("ar", "ar_EG", "arb_Arab", "acm_Arab", "xquad.ar", "20231101.ar", "ar-en").
+# "ace_Arab" is Acehnese written in Arabic script, so it deliberately does not match.
+ARABIC_CODES = {
+    "ar", "ara", "arb", "arz", "ary", "arq", "aeb", "apc", "ajp", "acm", "acq", "acw",
+    "acx", "adf", "aec", "afb", "ars", "ayl", "ayp", "ayh", "abv", "shu", "ssh", "pga",
+    "arabic",
+}
+
+
+def is_arabic_config(name: str) -> bool:
+    for tok in re.split(r"[^A-Za-z0-9]+", name or ""):
+        t = tok.lower()
+        if t in ARABIC_CODES or re.fullmatch(r"ar\d+", t):
+            return True
+    return False
 
 
 def get(path, **params):
     url = f"{API}/{path}?" + urllib.parse.urlencode(params)
-    try:
-        with urllib.request.urlopen(urllib.request.Request(url, headers=UA), timeout=60) as r:
-            return json.load(r)
-    except Exception as exc:
-        return {"__error__": f"{type(exc).__name__}: {exc}"}
+    for attempt in range(3):
+        try:
+            with urllib.request.urlopen(urllib.request.Request(url, headers=UA), timeout=60) as r:
+                return json.load(r)
+        except Exception as exc:
+            err = f"{type(exc).__name__}: {exc}"
+            if not any(t in err for t in ("429", "500", "502", "Timeout")) or attempt == 2:
+                return {"__error__": err}
+            time.sleep(4 * (attempt + 1))
 
 
 def first_split(dataset):
+    """The September 2026 rule. Kept for the record; do not use for new scans."""
     d = get("splits", dataset=dataset)
     if "__error__" in d or not d.get("splits"):
         return None, None, d.get("__error__", "no splits")
-    for s in d["splits"]:                      # prefer a train/test split named plainly
+    for s in d["splits"]:
         if s["split"] in ("train", "test"):
             return s["config"], s["split"], None
     s = d["splits"][0]
     return s["config"], s["split"], None
 
 
+def choose_split(dataset):
+    """Return (config, split, rule, error). Prefer an Arabic-coded config."""
+    d = get("splits", dataset=dataset)
+    if "__error__" in d or not d.get("splits"):
+        return None, None, None, d.get("__error__", "no splits")
+    splits = d["splits"]
+    configs = {s["config"] for s in splits}
+    arabic = [s for s in splits if is_arabic_config(s["config"])]
+    if arabic:
+        pool, rule = arabic, "arabic-config"
+    elif len(configs) == 1:
+        pool, rule = splits, "single-config"
+    else:
+        pool, rule = splits, "no-arabic-config-name"
+    for s in pool:
+        if s["split"] in ("train", "test"):
+            return s["config"], s["split"], rule, None
+    s = pool[0]
+    return s["config"], s["split"], rule, None
+
+
+def _strings(v):
+    """Every non-empty string inside a row value, including in lists and records."""
+    if isinstance(v, str):
+        if v.strip():
+            yield v
+    elif isinstance(v, dict):
+        for x in v.values():
+            yield from _strings(x)
+    elif isinstance(v, list):
+        for x in v:
+            yield from _strings(x)
+
+
+_ARABIC_LETTER = re.compile(r"[ء-ي]")
+
+
 def audit(dataset, n=100):
-    config, split, err = first_split(dataset)
+    config, split, rule, err = choose_split(dataset)
     if err:
         return {"dataset": dataset, "error": err}
     rows = get("rows", dataset=dataset, config=config, split=split, offset=0, length=n)
     if "__error__" in rows or "rows" not in rows:
         return {"dataset": dataset, "error": rows.get("__error__") or str(rows)[:120]}
 
-    total = bad = spans = 0
+    total = arabic = bad = spans = 0
+    severity = {}
     worst = None
     for r in rows["rows"]:
-        for v in r.get("row", {}).values():
-            if not isinstance(v, str) or not v.strip():
-                continue
+        for v in _strings(r.get("row", {})):
             total += 1
+            if _ARABIC_LETTER.search(v):
+                arabic += 1
             rep = scan_text(v)
             if rep.findings:
                 bad += 1
                 spans += len(rep.findings)
+                for f in rep.findings:
+                    sev = getattr(f, "severity", None)
+                    if sev:
+                        severity[sev] = severity.get(sev, 0) + 1
                 if worst is None:
                     worst = rep.findings[0].text[:60]
-    return {"dataset": dataset, "config": config, "split": split,
+    return {"dataset": dataset, "config": config, "split": split, "config_rule": rule,
             "rows_sampled": len(rows["rows"]), "text_fields": total,
-            "fields_corrupted": bad, "spans": spans, "example": worst,
+            "arabic_fields": arabic, "fields_corrupted": bad, "spans": spans,
+            "severity": severity or None, "example": worst,
             "pct": round(100 * bad / total, 2) if total else 0.0}
 
 
@@ -75,7 +155,8 @@ if __name__ == "__main__":
         if "error" in r:
             print(f"  {t:52} ERROR {r['error'][:60]}")
         else:
-            flag = "  <-- CORRUPTION" if r["fields_corrupted"] else ""
+            flag = "  <-- FINDINGS" if r["fields_corrupted"] else ""
             print(f"  {t:52} {r['fields_corrupted']:>4}/{r['text_fields']:<5} fields "
-                  f"({r['pct']:>5}%) in {r['rows_sampled']} rows{flag}")
+                  f"({r['pct']:>5}%), {r['arabic_fields']} Arabic, config "
+                  f"{r['config']} [{r['config_rule']}]{flag}")
     json.dump(out, open("audit.json", "w"), ensure_ascii=False, indent=2)

targets.json

@@ -0,0 +1,343 @@
+[
+ "2A2I/Arabic_Aya",
+ "Abdelkareem/Arabic-article-summarization-30-000",
+ "Abdelkareem/arabic-article-summarization",
+ "Abdelkareem/arabic-bbc-news",
+ "Abdelkareem/arabic_summarization_text",
+ "Abdelkareem/rwkv_Alpaca_arabic_instruct_version",
+ "Abdelkareem/simple-benchmark-arabic-summarization",
+ "Abdelkareem/wikihow-arabic-summarization",
+ "Abdelrahman-Rezk/Arabic_Dialect_Identification",
+ "Abdelrahman2922/Egyptian-Arabic-synthetic-stt",
+ "AdhamAshraf/egyptian-2-arabic",
+ "AhmadHakami/saudipedia-arabic-qa",
+ "AhmedBoin/Arabic-ASR",
+ "AhmedBou/Arabic_instruction_dataset_for_llm_ft",
+ "AhmedBou/MMMLU_arabic_law_Instruct",
+ "AhmedBou/MMMLU_arabic_medicine_Instruct",
+ "AhmedGaber77/ArabicNews",
+ "AhmedTaha012/arabic_text_yolov8_V0.5",
+ "Amineyyyhh/Arabic_Sentiment_Twitter_Corpus",
+ "Amr-khaled/Egyptian-Arabic_English_V1",
+ "ApexVOrteX-1/egyptian-arabic-style-dataset",
+ "ArabicNewsAnalyzer/Topic-Modeling-Checkpoints",
+ "ArabicSpeech/ArA-DF-2026",
+ "AuthenticIlm/Shamela4_Full_DB",
+ "Ayham7addad/Arabic_Formal_Economic_News_Jordan",
+ "Bisher/SadeedDiac-25_predictions_train_run-qwen2.5-1.5b-instruct-arabic-diacritization-moaaz-fadel_10k",
+ "BounharAbdelaziz/Arabic-Synthetic-Summarization-Dataset",
+ "BounharAbdelaziz/Arabic-Synthetic-Summarization-Dataset-Filtered",
+ "Buraaq/quran-md-ayahs",
+ "CNTXTAI0/arabic_dialects_question_and_answer",
+ "Chiplunkar/MMMLU_arabic_law_Instruct",
+ "ClusterlabAi/101_billion_arabic_words_dataset",
+ "CohereLabs/Global-MMLU",
+ "CohereLabs/Global-MMLU-Lite",
+ "CohereLabs/include-base-44",
+ "CohereLabs/xP3x",
+ "Davlan/sib200",
+ "Dr-AliGomaa/ar-quran-hadith14books-MSA",
+ "DrAbdulmalek/arabic-medical-ocr-corrections",
+ "Elfsong/Arabic_Dialect_DPO",
+ "Falah/classification_arabic_dialects",
+ "FaresElmenshawi/quran-recitations-asr",
+ "Fatimah8Moheeb/Arabic-Poetry-Dataset",
+ "FatimahEmadEldin/Gutenberg-Arabic-OCR-HTML-Pages",
+ "FatimahEmadEldin/alsanaa-emirati-arabic-asr",
+ "FatimahEmadEldin/deepseek-ocr-arabic-v1",
+ "FreedomIntelligence/Evol-Instruct-Arabic-GPT4",
+ "FreedomIntelligence/MMLU_Arabic",
+ "FreedomIntelligence/evol-instruct-arabic",
+ "GEM/wiki_lingua",
+ "GEM/xlsum",
+ "Gheras/Financial_Reasoning_QA_Arabic_Dataset",
+ "Gheras/Financial_Reasoning_QA_Arabic_Dataset_V2",
+ "GoAGI-AI/STEM_arabic_qa_1k",
+ "HPLT/DocHPLT",
+ "HPLT/HPLT2.0_cleaned",
+ "Haneen84/Arabic_news",
+ "Haneen84/Arabic_news_articles_Brexit",
+ "HaoElshazly/quran_data",
+ "HebArabNlpProject/ArabicSentimentDataSet",
+ "Helsinki-NLP/opus-100",
+ "HeshamHaroon/Arabic_fake_news_dataset",
+ "HeshamHaroon/QA_Arabic",
+ "HeshamHaroon/arabic-dialect-dpo",
+ "HeshamHaroon/arabic-turath-ocr",
+ "HuggingFaceFW/finepdfs-edu",
+ "HumynLabs/Arabic_Documents_Dataset_PDF",
+ "ISLAM-PO/arabic-history-and-dialects",
+ "ISLAM-PO/documents-Egyptian-Arabic",
+ "IbrahimAmin/egyptian-arabic-fake-reviews",
+ "IbrahimAmin/egyptian-arabic-hate-speech",
+ "IbrahimSalah/The_Arabic_News_speech_Corpus_Dataset",
+ "Iftoo95/Arabic_Sentiment_and_Topics",
+ "ImranzamanML/Arabic-Sentiments",
+ "InfoBayAI/Arabic-Non-STEM-QA-MCQ-Dataset",
+ "InfoBayAI/Arabic-STEM-QA-MCQ-Dataset",
+ "JayanthMuthu/arabic-ocr",
+ "KFUPM-JRCAI/arabic-acoustic-poetry-benchmark",
+ "KFUPM-JRCAI/arabic_dialects_dataset_experimental",
+ "Kamyar-zeinalipour/Arabic-Clue-Instruct",
+ "Kamyar-zeinalipour/arabic_multi_dialect_dialogue",
+ "Khaledtelbahnasy/egyptian-arabic-stt-data",
+ "LHF/escorpius-mr",
+ "M-A-D/Mixed-Arabic-Datasets-Repo",
+ "MBZUAI/ArabicMMLU",
+ "MBZUAI/Dialectal-Arabic-MMLU",
+ "MBZUAI/human_translated_arabic_mmlu",
+ "Madu786/Arabic_Sentiment_Dataset",
+ "Madu786/multilingual-urdu-romanurdu-arabic-english-sentiment",
+ "Mo-Abdalkader/Egyptian-Arabic-English-Parallel-Corpus",
+ "Mohamed-Sami/arabic-instruction-fine-tuning-prep",
+ "MohamedRashad/MASC-Arabic",
+ "MohamedRashad/arabic-books",
+ "MohamedRashad/arabic-img2md",
+ "MohammedNasser/Arabic_Reasoning_Instruct_QA",
+ "Monda/arabic_poetry",
+ "MustafaIbrahim/50k-Arabic-QA",
+ "MustafaIbrahim/medical-arabic-qa",
+ "Nahla-yasmine/arabic_fake_news",
+ "NajahUniv/arabic-univeristy-chatbot-qa",
+ "NightPrince/islamic-arabic-qa",
+ "Nooh/arabic-asr-audio",
+ "OALL/AlGhafa-Arabic-LLM-Benchmark-Native",
+ "OALL/Arabic_EXAMS",
+ "OALL/details_CohereForAI__c4ai-command-r7b-arabic-02-2025_v2",
+ "OALL/details_deep-analysis-research__D2IL-Arabic-Qwen2.5-72B-Instruct-v0.1_v2",
+ "OALL/details_deep-analysis-research__D2IL-Arabic-Qwen2.5-72B-Instruct-v0.1_v2_alrage",
+ "OALL/details_deep-analysis-research__D2IL-Arabic-Qwen2.5-72B-Instruct-v0.2_v2",
+ "OALL/details_deep-analysis-research__D2IL-Arabic-Qwen2.5-72B-Instruct-v0.2_v2_alrage",
+ "OALL/details_tunny__Arabic_Qwen2.5_72B_instruct_finetune_0.1_v2",
+ "Omar-youssef/English-Egyptian-Arabic-Translation",
+ "Omar-youssef/English-Egyptian-Arabic-Translation-Pairs",
+ "Omar-youssef/arabic-ocr-markdown-dataset",
+ "Omar-youssef/arabic-ocrr-dataset",
+ "Omar-youssef/islamic-qa-egyptian-arabic",
+ "Omartificial-Intelligence-Space/FineWeb2-Egyptian-Arabic",
+ "OpenAssistant/oasst1",
+ "OpenAssistant/oasst2",
+ "PRAli22/Arabic_dialects_to_MSA",
+ "Pangeanic/Iraqi-Arabic-multidomain-QA-text",
+ "PleIAs/common_corpus",
+ "PoetryMTEB/ArabicPoetryMeterClassification",
+ "PoetryMTEB/ArabicPoetryPeriodsClassification",
+ "PoetryMTEB/ArabicPoetryThemeClustering",
+ "QCRI/ImageEval-ArabicNLP26",
+ "QCRI/arabic_pos_dialect",
+ "Qanadil/ASTD_Arabic_Sentiment_Tweets_Dataset",
+ "Qanadil/ArSAS_An_Arabic_Speech-Act_and_Sentiment_Corpus_of_Tweets",
+ "Quran-Lab/QuranTTS",
+ "Quran-Lab/quran-tajweed-phonetics",
+ "RASHID778/arabic-instruct-curated",
+ "Rabe3/egyptian-arabic-tts-diacritized",
+ "ReejaNoor/arabic-food-asr-dataset",
+ "Sabri12blm/Arabic-Quran-ASR-dataset",
+ "Sadique5/arabic_quranic_asr",
+ "Sadique5/openslr_quranic_asr",
+ "SahmBenchmark/arabic-financial-qa_eval",
+ "SahmBenchmark/arabic-financial-qa_train",
+ "SaiedAlshahrani/Egyptian_Arabic_Wikipedia_20230101",
+ "Salama1429/tarteel-ai-everyayah-Quran",
+ "Salesteq/arabic-dialect-text",
+ "Salesteq/arabic-dialects-gold20",
+ "Salesteq/arabic-dialects-gold20-code-switch",
+ "Shabdobhedi/MMMLU_arabic_law_Instruct",
+ "Shabdobhedi/MMMLU_arabic_law_Instruct_New",
+ "Shaer-AI-2/shaer-eval-raw-gpt2-medium-arabic-poetry",
+ "Shaer-AI-2/shaer-eval-raw-gpt2-small-arabic-poetry",
+ "Skiittoo/aljazeera-news-arabic",
+ "SultanR/fineweb-edu-arabic",
+ "TheRealOKAI/arabic-ocr-datasetv3",
+ "TheRealOKAI/arabic_ocr_merged_dataset",
+ "TigreGotico/arabic-stem-lexicon",
+ "Umartoyin/Accented_arabic-asr-dataset",
+ "Umartoyin/arabic-asr-dataset",
+ "Unimatch-Ai/arabic_chartqa_qa_pseudo_v1",
+ "Walid-Ahmed/arabic_news",
+ "Yahya-Mohamed/egyptian-arabic-speech-dataset",
+ "Yasbok/Alpaca_arabic_instruct",
+ "Yousefmd/arabic_ocr_dataset",
+ "YoussefAnwar/Arabic-news",
+ "aallail/arabic_ocr_synth_2",
+ "aallail/arabic_ocr_synth_3",
+ "abdalrahmanshahrour/autotrain-data-auto-arabic-summarization",
+ "abdo1819/arabic-english-code-switching-synthetic-asr",
+ "abzoo/arabic-names-synthetic-ocr",
+ "ahishamm/QURANICWhisperDataset",
+ "ahmed220v/arabic_quranic_asr",
+ "ahmedafareed/Syl_arabicnewscorpus",
+ "ahmedafareed/arabicnewscorpus",
+ "ahmedafareed/emb_Syl_Class_arabicnewscorpus",
+ "ahmedheakl/arabic_arabicocr",
+ "ahmedheakl/arocrbench_arabicocr",
+ "akbargherbal/six_millions_instruction_dataset_for_arabic_llm_ft",
+ "alalfi/arabic-summarization",
+ "alalfi/arabic-summarization02",
+ "alexandrainst/m_arc",
+ "alexandrainst/m_hellaswag",
+ "alhasanmuhammad/Translated_Egyptian_Arabic_Lectures",
+ "allenai/c4",
+ "amgadhasan/arabic_tweets_dialects",
+ "aminedjebbie/Multi-Arabic-dialects",
+ "ammarthabet/egyptian-arabic-speech",
+ "arbml/Arabic_Dialects_Dataset",
+ "arbml/Arabic_News",
+ "arbml/Arabic_News_Tweets",
+ "arbml/Arabic_Sentiment_Twitter_Corpus",
+ "arbml/ultimate_arabic_news",
+ "arsaporta/symile-m3",
+ "artur-muratov/multilingual-speech-commands-15lang",
+ "asas-ai/Arabic-article-summarization",
+ "asas-ai/Arabic_Sentiment_Twitter_Corpus",
+ "badrex/MADIS5-spoken-arabic-dialects",
+ "bigscience/xP3",
+ "bigscience/xP3all",
+ "bigscience/xP3mt",
+ "bobez999/arabic-qa-dataset-sigir2024",
+ "cloudaocr/arabic-synthetic-scanned-books",
+ "craneset/arabic-ocr",
+ "dataflare/arabic-dialect-corpus",
+ "deep-analysis-research/details_D2IL-Arabic-Qwen2.5-72B-Instruct-v0.1",
+ "deepdml/Quran_Speech_Dataset",
+ "disco-eth/WorldSpeech",
+ "dispatchAI/arabic-poetry-instructions",
+ "drelhaj/Arabic-Dialects",
+ "drelhaj/Arabic-news-and-management-corpus",
+ "ebubekr53/organic-egyptian-arabic-dialect-dataset",
+ "ebubekr53/organic-gulf-arabic-dialect-dataset",
+ "ebubekr53/organic-levantine-arabic-dialect-dataset",
+ "ebubekr53/organic-maghrebi-arabic-dialect-dataset",
+ "ebubekr53/organic-sudanese-arabic-dialect-dataset",
+ "ehabnegm/100-hour-Egyptian-dataset-single-speaker",
+ "emadjumaah/mishkat-quran-audio",
+ "epfml/FineWeb2-HQ",
+ "epfml/FineWeb2-embedded",
+ "facebook/belebele",
+ "facebook/xnli",
+ "fr3on/arabic-dialect-corpus",
+ "freococo/ocr_arabic_books",
+ "freococo/synth_shamela_ocr_arabic_books",
+ "fsicoli/common_voice_15_0",
+ "fsicoli/common_voice_17_0",
+ "fsicoli/common_voice_22_0",
+ "gagan3012/arabic-news-clustering-p2p",
+ "geeeezx/egyptian-arabic-400k",
+ "gililior/mmlu-prox-eval-predictions",
+ "google-research-datasets/tydiqa",
+ "google/svq",
+ "google/wmt24pp",
+ "google/xquad",
+ "google/xtreme",
+ "gsarti/flores_101",
+ "haidermasood99/arabic_quran_asr_final",
+ "haidermasood99/quran_asr_final",
+ "hajerbchn/MMMLU_arabic_Medicine_Instruct",
+ "hajerbchn/MMMLU_arabic_law_Instruct",
+ "hajerbchn/arabic-medical-qa",
+ "hammoualiyoucef20/quran-audio",
+ "hastyle/arabic-manuscript-ocr",
+ "herwoww/arabic_xvector_embeddings",
+ "heshamelmahdi/egyptian-arabic-orpheus-tokenized",
+ "hetchyy/quranic-universal-ayahs",
+ "hltcoe/megawika",
+ "hmmamalrjoub/arabic-sentiments2",
+ "hozifa1/quran_and_sunnah",
+ "ieasybooks-org/prophet-mosque-library",
+ "ieasybooks-org/shamela-waqfeya-library",
+ "ieasybooks-org/waqfeya-library",
+ "ilyes25/ASR_arabic",
+ "ilyes25/ASR_arabic_modified2",
+ "ilyes25/Asr_arabic_modified",
+ "insanalamin/quran-terjemahan-indonesia-audio-by-ayah",
+ "ismaeeelxd/Egyptian-Arabic-Lectures",
+ "josephnashat/ArabicSentimentAnalysis",
+ "juletxara/xstory_cloze",
+ "kamalamarouche/MMMLU_arabic_law_Instruct",
+ "karimraouf/Arabic-Summarization-Dataset-AsDs",
+ "khalidalt/ultimate_arabic_news",
+ "kokojake/lmsys_egyptian_arabic_convs",
+ "kokojake/oasst2_egyptian_arabic_convs",
+ "legacy-datasets/common_voice",
+ "legacy-datasets/wikipedia",
+ "loay/arabic-ocr-synthetic-scans-faker-300k",
+ "madilcy/arabic-medical-qa-MERGED-MAQA-MMMLU-MI",
+ "mahmoudsaalama/arabic-eou-saudi-dialect",
+ "marait123/IBR_MMMLU_arabic_law_Instruct",
+ "marait123/IBR_MMMLU_arabic_medicine_Instruct",
+ "medmabfc/Arabic_News_Texts_Corpus",
+ "medyas/arabic-ocr-printed-500k",
+ "melsiddieg/qari-arabic-ocr-10k",
+ "mhardalov/exams",
+ "michsethowusu/english-egyptian-arabic_sentence-pairs_mt560",
+ "miscovery/General_Facts_in_English_Arabic_Egyptian_Arabic",
+ "mkay8/arabic_mental_health_QA",
+ "mksaad/Arabic_news",
+ "moazeldegwy/Arabic-Nutrition-QA",
+ "mohajesmaeili/Persian_Arabic_TextLine_Image_Ocr_Medium",
+ "mohajesmaeili/Persian_Arabic_TextLine_Image_Ocr_Small",
+ "moqarashad/arabic_corpus_international_news",
+ "mostafa-rabie/Arabic-ASR-Features",
+ "mustaphaelkady/arabic-ocr-books",
+ "nabbahi/traditionnals_arabic_shoes_split",
+ "nazimali/quran",
+ "nehalelkaref/classical_arabic_poetry",
+ "niamaelkhbir/Arabic_Dialect_NER",
+ "nizarun/Arabic-Poetry-Scansion",
+ "occiglot/tokenizer-wiki-bench",
+ "oddadmix/Qari-OCR-0.2.2-Arabic-2B_Qari-0.2-eval-tashkil",
+ "oddadmix/arabic-audio-collection-algerian-loubna-stories",
+ "oddadmix/arabic-news-summarization",
+ "oddadmix/dialectal-arabic-lahgtna-v2",
+ "oddadmix/egyptian-arabic-english-translation-50k",
+ "openai/MMMLU",
+ "openfoodfacts/product-database",
+ "openlanguagedata/flores_plus",
+ "pedramaa/Egyptian-Arabic-Data-LLM",
+ "projectqhaiderhabeeb/arabic-letters-asr",
+ "pyschopoodle/MMMLU_arabic_medicine_Instruct",
+ "qandos0/AFRD_Arabic-Fake-Reviews-Detection",
+ "quranlab/hadith",
+ "quranlab/quran",
+ "quranlab/quran-audio",
+ "quranlab/quran-tafsir",
+ "rabah2026/Quran-Ayah-Corpus",
+ "ramyibrahim/synthetic_egyptian-arabic-tts",
+ "rehabyehia/rev3_9463_tv_news_arabic",
+ "riotu-lab/ArabicQA_2.1M",
+ "sadeem-ai/arabic-qna",
+ "sadra-barikbin/crcis-quranic-eval-leaderboard-results_details_mistralai__Mistral-7B-v0.1_private",
+ "saeedbark/hadrami-arabic-dialect-dataset",
+ "saleh-c4/arabic-ocr-images",
+ "saleh-c4/arabic-ocr-labels",
+ "sanaa-11/fake-real-arabic-news",
+ "sarulab-speech/yodas2_sidon",
+ "sedra-hugface/arabic-handwritten-ocr-eval",
+ "sepidmnorozy/Arabic_sentiment",
+ "sherif1313/Historical-Arabic-Handwritten-OCR",
+ "stanford-oval/ccnews",
+ "styletts2-community/multilingual-pl-bert",
+ "tahaalselwii/arabic-poetry-verses-and-explanations",
+ "tahaio/arabicpoetry",
+ "tamm5y5m5/quran_nemotron_dual_v5_npl",
+ "tarekys5/Arabic_Healthcare_QA",
+ "thethanksforthegod/quran-asr-mega-corpus",
+ "tiiuae/Falcon-Arabic-7B-Instruct-details",
+ "unimelb-nlp/wikiann",
+ "unohamza/Arabic-news-daily",
+ "uonlp/CulturaX",
+ "wala0404/ARABIC-NEWS",
+ "wikimedia/wikipedia",
+ "wissamantoun/qasr-arabic-speech-continuations",
+ "xtz999/egyptian_arabic_speech_zaid",
+ "yallashoot/arabic-fan-sentiment",
+ "yallashoot/arabic-transfer-news",
+ "yrrhall/arabic-dialect-corpus",
+ "zafro77/Arabic-Sentiments",
+ "zaibihassan/Quranic-Recitation-Data",
+ "zaibihassan/Quranic-Translation-Audio-Data",
+ "zaibihassan/Quranic-Word-By-Word-Audio-Data",
+ "zainulhakim/Arabic_ASR",
+ "zainulhakim/Stella20_Arabic_ASR",
+ "zainulhakim/Stella_Arabic_ASR"
+]