mbart-large-50-many-to-many-mmt-finetuned-fij_Latn-to-eng_Latn

This model is a fine-tuned version of facebook/nllb-200-distilled-600M on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.9598
• Bleu: 45.0972
• Gen Len: 42.752

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 6
• total_train_batch_size: 12
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
2.4054	0.49	500	1.7028	24.9597	43.04
1.6855	0.98	1000	1.3701	33.3128	42.2
1.4042	1.47	1500	1.2224	37.6016	43.536
1.2991	1.96	2000	1.1467	40.3541	42.428
1.1819	2.45	2500	1.0950	42.2106	42.58
1.1323	2.94	3000	1.0523	42.9418	42.76
1.0676	3.43	3500	1.0238	43.4974	42.684
1.0404	3.93	4000	1.0082	43.6092	42.616
0.9882	4.42	4500	0.9942	44.7199	42.912
0.982	4.91	5000	0.9814	44.8061	42.516
0.9372	5.4	5500	0.9781	44.3808	42.476
0.9382	5.89	6000	0.9675	45.0267	42.76
0.915	6.38	6500	0.9659	45.0073	42.676
0.9126	6.87	7000	0.9617	44.9582	42.548
0.8903	7.36	7500	0.9609	44.8713	42.724
0.8873	7.85	8000	0.9598	45.0972	42.752

Framework versions

• Transformers 4.21.3
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1