Submit final outputs through the official CodaBench competition pages.
Use the provided IDs exactly as released in the blind test files.
Do not manually inspect or infer hidden gold labels during the evaluation phase.
Teams may submit to one or more subtasks and tracks.
Final rankings are produced only from official submissions made before the deadline.
Subtask 1 Output
Provide one dialectal Arabic translation for each required English turn.
Preserve the source ID, target dialect or country label, and predicted translation field.
Constrained submissions must use only released Alexandria training and development data.
Constrained models must not exceed 5B total parameters.
spBLEU is the primary metric; chrF++ is also reported overall and per country.
Subtask 2 Output
TBD
Subtask 3 Output
Provide predicted word-level error spans in the translated text.
Assign one error category from the official typology to each span: graphetics, morphosyntax,
orthography_writing_conventions, pragmatics, semantics, or sociolinguistics.
Use the span-indexing convention specified in the evaluation scripts.
Empty predictions should be represented using the official no-error format.
Outputs are ranked by Overall Score, computed as the average of Exact Match F1 and Error Class
Macro-F1, with additional per-direction and category breakdowns.
Verification Checklist
Run the released validation script before uploading predictions.
Confirm UTF-8 encoding and valid JSON/CSV formatting.
Check that every test example has exactly the expected number of predictions.
Keep a copy of the submitted system outputs and configuration for the paper.
Email the organizers before the deadline if the platform rejects a valid file.
