abdelhaqueidali/Amazigh-Facebook-Data-Export

Dataset Card for Amazigh Facebook Dataset (Filtered)

This dataset contains a curated, filtered collection of personal Facebook posts and comments written by an Amazigh speaker. It captures real-world communication usage of Amazigh in Tifinagh, the speaker used Southern Moroccan Amazigh (Tashelhit) and leans to the extent possible to Standard Moroccan Amazigh using the Amazigh script (Tifinagh). Some posts and comments uniquely contains parallel multi-script and multi-languages (transliterations and translations embedded within some blocks) as well as grouped iterative edit history logs (drafts).

Dataset Details

Dataset Description

The Amazigh Facebook Dataset is an organic linguistic corpus built from direct social media exports. Communication on social media platforms is fluid; while this dataset is predominantly written in the Amazigh script (Tifinagh), it naturally exhibits multi-script characteristics. Certain entries include parallel Latin transliterations and English/Arabic context translations.

To maintain high linguistic relevance, a strict density scanner was utilized to drop any text log where Arabic script characters made up 90% or more of the total alphabetic character volume as well as when Tifinagh is less than 10% so most of where Amazigh is present is what is Available.

• Curated by: Abdelhaque Id Ali (Facebook Profile Owner)
• Language(s) (NLP): Southern Moroccan Amazigh (Tashelhit) and Standard Moroccan Amazigh written primarily in Tifinagh script; features parallel Latin and Arabic transliterations, as well as English and Arabic translations in some cases.
• License: CC-BY-4.0

Uses

Direct Use

• Language Modeling: Training and fine-tuning LLMs on organic, digital-native Amazigh text structures.
• Script Alignment & Machine Translation: Utilizing internal parallel blocks to map Tifinagh text to its Latin script transliterations and translations.
• Grammatical Error Correction (GEC): Training models to learn how real users revise, adjust, and correct spelling or syntax layout by analyzing the _drafts sequence files.

Out-of-Scope Use

This dataset is not intended for out-of-context linguistic normalization studies, as it features colloquial digital speech rather than strictly formal standard prose. It should not be used to identify or deanonymize any participating profile entities.

Dataset Structure

The final shared dataset files are organized into two distinct directory layers matching their system entry types:

📁 dataset_root/
  ├── 📄 comments.json
  ├── 📄 comments_drafts.json
  ├── 📄 posts.json
  └── 📄 posts_drafts.json

1. Clean Files Schema (comments.json, posts.json)

Flat logs containing a singular text entry per unique id.

[
  {
    "id": 847291043,
    "text": "ⵜⵉⴼⴰⵡⵉⵏ! ⵉⵙ ⵣⴰ ⵖⵓⵔⵓⵏ ⵔⵖⴰⵏ ⵡⵓⵙⵙⴰⵏ ⴰⴷ?"
  }
]

2. Draft Files Schema (comments_edits_drafts.json, posts_edits_drafts.json)

Hierarchical logs tracking revision evolution. draft_final represents the ultimate published state, while draft_1 represents the oldest tracked modification.

[
  {
    "id": 921504888,
    "draft_final": "ⴰⵣⵓⵍ ⵏⵏⵓⵏ ⴰⵢⵜⵎⴰ.",
    "draft_2": "ⴰⵣⵓⵍ ⵏⵏⵓⵏ ⴰⵢⵜⵍⴰ.",
    "draft_1": "ⴰⵣⵓⵍ ⴼⵍⵍⴰⵡⵏ ⴰⵢⵜⵍⴰ."
  }
]

Field Descriptions

Field Name	Data Type	Description
id	Integer	A unique, randomly assigned 9-digit numerical sequence key tracking the text entry.
text	String	Found exclusively in posts.json and comments.json; represents raw sanitized text string values.
draft_final	String	Found exclusively in _drafts.json configurations; represents the final state of an edit sequence.
draft_X	String	Found exclusively in _drafts.json configurations; represents ancestral historical edits leading up to finality.

Dataset Creation

Curation Rationale

Digital datasets for the Amazigh language—specifically utilizing the Tifinagh script—remain sparse. This dataset provides authentic, non-synthetic conversational records that reflect modern internet usage, script adaptations, and natural editing habits of Amazighs online.

Source Data

Data Collection and Processing

• Extraction: Raw textual data was extracted locally from standard Facebook data export dumps (.json).
• Anonymization: A custom parsing engine scanned all comment layers. Explicit user reference handles (mentions) were removed and replaced globally with a privacy placeholder ([@removed_name]), though embedded string names mentioned are not removed for technical limitations (no wat to identify them easily).
• Draft Alignment: For edit files, strings were processed through a sequence matcher. Text variants matching at a 55% similarity threshold were grouped under a single shared id to build the revision history profiles.
• Language Filtering: Script density analysis was applied to every record. Characters were analyzed by Unicode ranges. If the Arabic character block (\u0600-\u06ff) comprised 75% or more of the text block's raw letter count, the entire entry was discarded.

Who are the source data producers?

The data was generated organically by an Amazigh content creator and language enthusiast interacting on social media (Abdelhaque Id Ali).

Personal and Sensitive Information

All platform usernames, tags, and tracking indicators within the comments dataset were stripped and randomized via the automated cleaning pipeline, but embedded names mentions were not possible to be removed because of the technical limitations. Other metadata are removed.

Bias, Risks, and Limitations

• Dialectal Focus: The vocabulary and phrasing primarily mirror standard Moroccan Amazigh variations and localized regional expressions.
• Conversational Tone: The syntax represents casual digital messaging, meaning it features shorter sentence lengths, platform abbreviations, and exclamation conventions not typically present in textbook literature.

Glossary

• Tifinagh: The writing script used by Amazighs to write the Amazigh language.
• Drafts Clustering: The structural group mapping where individual modifications to a single post are captured sequentially rather than treated as unrelated data records.

Dataset Card Authors

Abdelhaque Id Ali: abdelhaque.idal@gmail.com