LmaanaDataset

Cleaned transcription data derived from the transcribed 100-gt-2.5 subset of atlasia/MoulSot-Full.

Contents

• 81,603 non-empty transcription rows.
• original_text: source transcription.
• new_transcription: cleaned transcription with selected French loanwords restored to Latin script.
• training_text: recommended text field for training.
• review_status: indicates whether a candidate conversion was applied.
• audio/100-gt-2.5/: original audio-bearing Parquet shards from MoulSot-Full.

Cleaning

• Removed 247 empty transcription rows.
• Removed no duplicate IDs.
• Normalized whitespace and line breaks.
• Preserved existing Latin-script code-switching.
• Converted only a conservative list of likely French loanwords written in Arabic script.

The audio-bearing Parquet shards are included in this repository under audio/100-gt-2.5/. Join their id and audio fields with moulsot_cleaned_for_training.csv; use training_text as the corrected target. Candidate conversions should be checked against audio before treating them as definitive ground truth.

Source

Atlasia, MoulSot-Full, 100-gt-2.5 configuration. The source dataset contains Moroccan Darija speech with natural French and English code-switching.