NLLB-600m-vie_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: 1.1189
• Bleu: 36.6767
• Gen Len: 47.504

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: 3
• eval_batch_size: 3
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 24
• 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
1.9294	2.24	1000	1.5970	23.6201	48.1
1.4	4.47	2000	1.3216	28.9526	45.156
1.2071	6.71	3000	1.2245	32.5538	46.576
1.0893	8.95	4000	1.1720	34.265	46.052
1.0064	11.19	5000	1.1497	34.9249	46.508
0.9562	13.42	6000	1.1331	36.4619	47.244
0.9183	15.66	7000	1.1247	36.4723	47.26
0.8858	17.9	8000	1.1198	36.7058	47.376
0.8651	20.13	9000	1.1201	36.7897	47.496
0.8546	22.37	10000	1.1189	36.6767	47.504

Framework versions

• Transformers 4.22.1
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1