Datasets:

---

syamjithnk

/

arabic-corpus-audit

Like 0

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.ai/code/session_01GaKH7wkxcBhgmbNVzYaxR2