Ammar-06/mladi-marbert-r07

R07-marbert: multi-label Arabic dialect identification

Fine-tuned from UBC-NLP/MARBERT for multi-label country-level dialect identification (18 sigmoid outputs, one per country). Bachelor thesis project, German International University.

• Training data: NADI 2020/2021/2023 training tweets with the LahjatBERT multi-label labels (56,113 unique texts, file nadi_lahjatbert_card1to18.jsonl; links and mentions replaced, no other normalization).
• Recipe: lower 8 encoder layers frozen, dropout 0.3, AdamW lr 5e-05, 500 warmup steps, batch 24, 2 epochs, fp16, max length 128, seed 42; epoch chosen by macro F1 on a 10% validation split of the training data.
• Use: sigmoid over the 18 logits, a dialect is valid if its probability is >= 0.3 (MLADI leaderboard method predict_binary_outcomes).
• Label order: Algeria, Bahrain, Egypt, Iraq, Jordan, Kuwait, Lebanon, Libya, Morocco, Oman, Palestine, Qatar, Saudi_Arabia, Sudan, Syria, Tunisia, UAE, Yemen.
• MLADI dev set (120 sentences, 8 dialects): macro F1 72.96, precision 74.40, recall 73.23.
• Training run card1to18_real_only_seed42, code commit c11a0b84f7c6eb8042bbd06b9a01a099dea40645.