FSTM-LIM/MoLeAd-A-Moroccan-Legal-Announcements-Dataset

MoLeAd: A Moroccan Legal Announcements Dataset

Dataset Link: https://huggingface.co/datasets/FSTM-LIM/MoLeAd-A-Moroccan-Legal-Announcements-Dataset

1. Dataset Description

MoLeAd (Moroccan Legal Announcements Dataset) is a curated dataset of official Moroccan legal announcements (bulletins officiels).
It is intended for NLP research in Arabic, including legal text classification, named entity recognition, and legal information extraction.

• Language: Arabic (Morocco)
• Domain: Legal / Official announcements
• Size: 483,262 announcements (~1.7 GB)
• Source: Moroccan official gazettes ("Bulletin Officiel")
• Content per announcement: company name, legal form, subject, body text, registration info, features, full text

---

2. Dataset Structure

Each record contains:

Field	Description
id	Unique identifier of the announcement
meta	Metadata about the announcement
meta.bo_number	Bulletin number
meta.bo_date	Publication date
meta.announcement_id	Announcement ID in the bulletin
meta.page	Page number in the bulletin
text_content	Core text information
text_content.company_name	Company / entity name
text_content.subject_raw	Original subject line
text_content.body	Body text of the announcement
text_content.full_text	Concatenated full text (includes company, body, features)
text_content.features	List of extracted features (legal form, capital, address, RC, etc.)
category	High-level classification of announcement (Creation or Modification)
subject_canonical	Canonicalized subject of the announcement
annotation	Annotation metadata
annotation.label_source	Source of annotation (manual, Gemini 3 Pro, bert-base-arabertv2)
annotation.gold_standard	Whether the annotation is manually verified
annotation.confidence_score	Confidence score (0–1) for AI annotations
annotation.agreement	Inter-annotator agreement (if applicable)

---

3. Annotations

The dataset includes three types of annotations:

1. Manual (Gold): 1,002 manually verified announcements
2. Gemini 3 Pro: ~50,000 high-confidence annotations
3. BERT-base-Arabertv2: Remaining ~432,260 announcements annotated by a fine-tuned AraBERT model

Normalization:

• Confidence scores were mapped from qualitative labels ("high", "medium", "low") to numerical values (0.9, 0.6, 0.3).
• The agreement field indicates inter-annotator consistency for manually annotated samples.

---

4. Recommended Usage

• Classification: Predict category or subject_canonical from raw text.
• Information extraction: Extract legal forms, partners, capital, and registration details.
• Evaluation / Benchmarking: Use the gold set (MoLeAd-A-Moroccan-Legal-Announcements-Dataset-gold) for testing and validation.
• Training / Fine-tuning: Use the weakly annotated portion for model training, streaming is recommended for large-scale usage.

---

5. Dataset Splits

Split	Samples	Description
Train	458,147	Weakly annotated for model training
Validation	24,113	Weakly annotated for model selection
Test (Gold)	1,002	Manually verified announcements for evaluation

Sharding: Dataset is split into shards (max 500MB each) for memory-efficient loading.

Example of loading:

from datasets import load_dataset

# Full dataset (streaming)
dataset = load_dataset(
    "FSTM-LIM/MoLeAd-A-Moroccan-Legal-Announcements-Dataset",
    split="train",
    streaming=True
)

# Gold set
gold_dataset = load_dataset(
    "FSTM-LIM/MoLeAd-A-Moroccan-Legal-Announcements-Dataset-gold",
    split="train"
)

---

8. Dataset Statistics

• Total samples: 483,262
• Gold annotations (manually verified): 1,000 (0.21%)
• Weak annotations (auto-labeled): 482,260 (99.79%)

Categories

Category	Count	Percentage
Modification	248,686	51.46%
Creation	234,576	48.54%

Top 10 Subjects

Subject Canonical	Count	Percentage
تأسيس شركة ذات المسؤولية المحدودة ذات الشريك الوحيد	137,831	28.52%
تأسيس شركة ذات المسؤولية المحدودة	89,935	18.61%
حل شركة	50,680	10.49%
إعلان متعدد القرارات	43,094	8.92%
تفويت الحصص الاجتماعية	32,188	6.66%
تحويل المقر الاجتماعي للشركة	23,744	4.91%
قفل التصفية	23,694	4.90%
رفع رأسمال الشركة	21,050	4.36%
تعيين مسير جديد شخص ذاتي أو اعتباري	15,788	3.27%
توسيع نشاط الشركة	6,402	1.32%

---

6. Dataset Card Notes

• License: Open Access / Academic Use (CC BY 4.0 suggested)
• Citation: Please cite the dataset in any research using it:** 1,002 manually verified announcements @dataset{MoLeAd-A, author = {FSTM-LIM}, title = {MoLeAd-A: Moroccan Legal Announcements Dataset}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/FSTM-LIM/MoLeAd-A-Moroccan-Legal-Announcements-Dataset}} }
• Ethical Considerations: The dataset contains publicly published legal information. Use responsibly.