run1

This model is a fine-tuned version of Helsinki-NLP/opus-mt-es-es on an unkown dataset. It achieves the following results on the evaluation set:

• Loss: 3.1740
• Bleu: 8.4217
• Gen Len: 15.9457

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: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
No log	1.0	250	4.2342	0.8889	83.4022
4.6818	2.0	500	3.7009	4.1671	35.587
4.6818	3.0	750	3.4737	7.6414	23.9674
3.4911	4.0	1000	3.3713	7.7512	18.6957
3.4911	5.0	1250	3.2689	8.0901	19.4674
3.0164	6.0	1500	3.2194	8.5708	25.0543
3.0164	7.0	1750	3.1853	9.5275	23.9239
2.6954	8.0	2000	3.1562	8.5635	18.9674
2.6954	9.0	2250	3.1564	8.2031	17.5978
2.4503	10.0	2500	3.1314	8.5638	18.1522
2.4503	11.0	2750	3.1511	8.8428	17.913
2.2554	12.0	3000	3.1513	8.1244	17.0
2.2554	13.0	3250	3.1664	8.0157	16.2717
2.1202	14.0	3500	3.1656	8.7758	16.6087
2.1202	15.0	3750	3.1550	8.4637	16.4565
2.0082	16.0	4000	3.1702	8.2488	15.8587
2.0082	17.0	4250	3.1725	8.609	16.3043
1.9274	18.0	4500	3.1750	8.4476	15.8043
1.9274	19.0	4750	3.1734	8.4753	16.5543
1.888	20.0	5000	3.1740	8.4217	15.9457

Framework versions

• Transformers 4.9.2
• Pytorch 1.9.0+cu102
• Datasets 1.11.1.dev0
• Tokenizers 0.10.3