NLLB-600m-nlg_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: 0.9402
• Bleu: 45.9717
• Gen Len: 42.476

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: 10000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
2.5032	0.49	500	1.7451	24.369	42.66
1.732	0.98	1000	1.3896	31.9939	42.304
1.4344	1.47	1500	1.2333	36.4344	42.384
1.3141	1.96	2000	1.1442	38.5023	41.96
1.1877	2.45	2500	1.0936	41.3292	42.668
1.1355	2.94	3000	1.0460	43.1357	43.22
1.0623	3.43	3500	1.0197	43.2339	42.692
1.0353	3.93	4000	1.0010	43.8863	43.012
0.9786	4.42	4500	0.9899	44.2478	43.012
0.9682	4.91	5000	0.9731	44.9191	42.816
0.9184	5.4	5500	0.9690	44.908	42.496
0.9208	5.89	6000	0.9558	45.5488	42.772
0.8854	6.38	6500	0.9561	45.7261	42.844
0.8815	6.87	7000	0.9495	45.1231	42.38
0.8543	7.36	7500	0.9475	45.6717	42.56
0.8462	7.85	8000	0.9442	45.9782	42.652
0.8422	8.34	8500	0.9436	45.9353	42.628
0.8323	8.83	9000	0.9407	45.7945	42.492
0.8218	9.32	9500	0.9405	46.0215	42.472
0.8226	9.81	10000	0.9402	45.9717	42.476

Framework versions

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