PontifexMaximus/TurkishTranslator

opus-mt-tr-en-finetuned-tr-to-en

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

• Loss: 0.6924
• Bleu: 54.7617
• Gen Len: 13.5501

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-06
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 16
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
No log	1.0	412	1.1776	43.3104	12.9297
1.4032	2.0	824	1.0750	45.7912	12.9155
1.2268	3.0	1236	1.0019	47.6255	12.9251
1.141	4.0	1648	0.9411	49.0649	12.9302
1.0651	5.0	2060	0.8929	50.4894	12.9066
1.0651	6.0	2472	0.8519	51.5072	12.9067
1.0025	7.0	2884	0.8180	52.5035	12.8875
0.9582	8.0	3296	0.7893	51.7587	13.5338
0.9173	9.0	3708	0.7655	52.3566	13.5376
0.8892	10.0	4120	0.7449	53.0488	13.5545
0.8639	11.0	4532	0.7285	53.5965	13.5539
0.8639	12.0	4944	0.7152	53.9433	13.5547
0.8424	13.0	5356	0.7053	54.2509	13.5502
0.8317	14.0	5768	0.6981	54.5339	13.5502
0.817	15.0	6180	0.6938	54.7068	13.5448
0.8155	16.0	6592	0.6924	54.7617	13.5501

Framework versions

• Transformers 4.19.2
• Pytorch 1.7.1+cu110
• Datasets 2.2.2
• Tokenizers 0.12.1