mbart_coded_sanlish

Fine-tuned facebook/mbart-large-50-many-to-many-mmt for coded-Sanlish → Bangla translation, part of the Santali Speech-to-Bangla-Text thesis pipeline.

• Source column: coded_sanlish (tagged as en_XX, a Latin-script placeholder)
• Target column: Bangla (tagged as bn_IN)
• Status: training complete — this is the best checkpoint

Training arguments used

Argument	Value
evaluation_strategy	epoch
save_strategy	epoch
logging_steps	20
learning_rate	0.0003
weight_decay	0.01
warmup_ratio	0.1
per_device_train_batch_size	8
per_device_eval_batch_size	8
gradient_accumulation_steps	2
num_train_epochs	8
predict_with_generate	True
fp16	True
load_best_model_at_end	True
metric_for_best_model	eval_loss
greater_is_better	False
save_total_limit	2

Final results (best checkpoint, by eval_loss)

Metric	Value
eval_loss	2.631115674972534
BLEU	7.6122363636138815
chrF++	25.5530785618982