---
language:
- kab
- ber
license: cc-by-2.0
tags:
- berber
- amazigh
- tamazight
- kabyle
- tifinagh
- transliteration
- tatoeba
configs:
- config_name: kab_tifinagh
  data_files:
  - split: train
    path: kab_tifinagh.tsv
- config_name: en_kab_tifinagh
  data_files:
  - split: train
    path: en_kab_tifinagh.tsv
- config_name: fr_kab_tifinagh
  data_files:
  - split: train
    path: fr_kab_tifinagh.tsv
task_categories:
- translation
- text-generation
---

# Dataset Card for Kabyle Latin-to-Tifinagh Parallel Corpus

This dataset provides a parallel corpus of the Kabyle language (Taqbaylit), pairing native Latin-based orthography with automated, context-aware Amazigh script transliterations. It is built by processing raw text data through a rule-based algorithmic pipeline designed to enforce strict orthographic purity, manage contextual phonetic mutations, and isolate foreign vocabulary.

## Dataset Details

### Dataset Description

- **Curated by:** Abdelhaque Id Ali
- **Language(s) (NLP):** Kabyle (Latin script and Amazigh script)
- **License:** CC-BY 2.0 (Inherited from the Tatoeba source corpus)

### Dataset Sources

- **Repository:** [More Information Needed]
- **Source Project:** Tatoeba Translation Corpus

## Uses

### Direct Use

- Training and fine-tuning machine translation models and sequence-to-sequence tokenizers.
- Evaluating local-first or low-resource script normalizers and converters.

### Out-of-Scope Use

This dataset is strictly optimized for native phonetic structures. Text processing forcefully filters out unadapted loanwords, making it unsuited for generalized code-switching (e.g., raw French-Kabyle mixed text) that does not adhere to standard Amazigh structural frameworks.

## Dataset Structure

The dataset contains three distinct configuration subsets formatted as Tab-Separated Values (`.tsv`) files with specific schemas:

### 1. Monolingual Subset (`kab_tifinagh`)
- `id`: Unique identifier inherited from the Tatoeba database.
- `kab_tfng`: The processed Amazigh alphabet transliteration column.
- `kab_latn`: The original source Kabyle sentence in standard Latin script.

### 2. English Parallel Subset (`en_kab_tifinagh`)
- `id`: Unique sentence identifier.
- `kab_tfng`: The processed Amazigh alphabet transliteration column.
- `kab_latn`: The source Kabyle sentence in standard Latin script.
- `en`: Corresponding English translation text string.

### 3. French Parallel Subset (`fr_kab_tifinagh`)
- `id`: Unique sentence identifier.
- `kab_tfng`: The processed Amazigh alphabet transliteration column.
- `kab_latn`: The source Kabyle sentence in standard Latin script.
- `fr`: Corresponding French translation text string.

## Dataset Creation

### Curation Rationale

While the Kabyle language is widely written using a modified Latin alphabet, the standardized Amazigh alphabet holds deep official and cultural significance. This dataset bridges the script divide algorithmically, applying rigorous phonetic harmony and morphological adjustments to build a high-fidelity parallel repository.

### Source Data

#### Data Collection and Processing

The source texts are extracted from the Kabyle subsets of the Tatoeba project. The pipeline maps characters directly to the **33 official IRCAM standardized alphabet characters** while executing a multi-tier linguistic and structural engine:

##### 1. Character Map & Realignment Graphemes
The pipeline uses a strict character matrix reducing entries directly to the core script block, explicitly accommodating gemination, labialization, and target replacements:
* **Geminated Affricates:** `čč`  =>  `ⵜⵜⵛ`, `ǧǧ`  =>  `ⴷⴷⵊ`, `žž`  =>  `ⴷⴷⵣ`, `ţţ`  =>  `ⵜⵜ`.
* **Labialization Graphemes:** `gʷ`  =>  `ⴳⵯ`, `kʷ`  =>  `ⴽⵯ`.
* **Standard Alternations:** `č`  =>  `ⵜⵛ`, `ǧ`  =>  `ⴷⵊ`, `ž`  =>  `ⴷⵣ`, `ţ`  =>  `ⵜⵜ`.
* **Foreign Borrowing Adaptations:** Maps validated morphological structures containing `p`  =>  `ⴱ` and `v`  =>  `ⴼ`.

##### 2. Grammatical Dash Merging
Structural hyphens used for direct/indirect objects, prepositions, or kinship suffixes are automatically stripped, merging dependent particles cleanly into single word tokens:
* **Prepositional/Object Frameworks:** `fell-as`  =>  `ⴼⴻⵍⵍⴰⵙ`, `ger-aɣ`  =>  `ⴳⴻⵔⴰⵖ`.
* **Locative/Locational Frameworks:** `ɣur-sen`  =>  `ⵖⵓⵔⵙⴻⵏ`, `deg-sen`  =>  `ⴷⴻⴳⵙⴻⵏ`.
* **Kinship Expressions:** `baba-s`  =>  `ⴱⴰⴱⴰⵙ`, `mmi-k`  =>  `ⵎⵎⵉⴽ`.
* *Note: Non-grammatical structural dashes are globally normalized to empty spaces.*

##### 3. Contextual Phonetics & Schwa (ⴻ) Optimization
* **Initial Mutations:** Word-initial sequences starting with `ye` or `we` migrate to native vocalic variants `i` and `u` respectively.
* **Vocalic Semivowel Linearizations:** Sequences matching `ew` or `ey` followed by a native consonant or sitting at the absolute end of a word are flattened directly into `u` or `i`.
* **Pharyngealization Harmony:** Progressive and regressive assimilation constraints force the plain alveolar trill `r` / `R` to upgrade into its pharyngealized counterpart `ṛ` / `Ṛ` if it directly borders or is separated by a single vowel from any pharyngeal anchor (`ẓ`, `ḍ`, `ṭ`, `ṣ`).
* **Schwa Stripping:** The engine drops an initial schwa (`ⴻ`) at the absolute start of a word. Internally, `ⴻ` is dynamically dropped between two distinct consonants using a strict regex lookahead filter, but remains preserved when sandwiched between identical, geminated consonants.

##### 4. Strict Isolation Filter (Skipped Sentences)
To prevent phonetic distortion, a row is completely discarded and sent to a `_skipped.tsv` ledger if a token meets any of the following conditions:
* Contains French accented characters (`é`, `è`, `â`, `ç`, etc.).
* Contains true diphthongs (doubled vowel sequences like `ae`, `ui`, `oi`).
* Contains the vowel `o` or `O` unless it represents a structural character assignment or explicit native override.
* Terminates explicitly with an unmapped trailing `e` or `E`.
* Contains an apostrophe (`'`) or represents an all-caps acronym (e.g., `MRI`).
* Contains standalone capital letters internally or at the tail end of strings.
* Fails a final structural purity regex check (`^[0-9ⵯ\u2D30-\u2D6F\s.,!?;:()""\']+$`).

##### 5. Target Overrides
High-frequency cross-linguistic tokens are explicitly caught by a manual override filter before any exclusion rules trigger:
* `Tom` / `ṭom` / `ṭum`  =>  `ⵟⵓⵎ`
* `Ait` / `ait`  =>  `ⴰⵢⵜ`
* `Mary`  =>  `ⵎⴰⵔⵉ`

#### Who are the source data producers?

The underlying source corpus sentences are crowdsourced, verified, and translated by independent contributors and native speakers on the open-source Tatoeba platform.

## Bias, Risks, and Limitations

Because the translation relies heavily on an automated algorithmic pipeline to catch external loanword structures, structural edge cases can occur:
- Loanwords that naturally lack explicit French accented markers, trailing `e`s, or common vowels like `o` may slip past the isolation filters and undergo incorrect, literal transliteration into the Amazigh script.
- Rare native structures or highly localized dialectal expressions matching an exclusion rule rule profile might be inadvertently skipped.