opus-mt-iir-en-finetuned-fa-to-en

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

• Loss: 1.0968
• Bleu: 36.687
• Gen Len: 16.039

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: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
3.1614	1.0	1509	2.8058	12.326	16.5467
2.7235	2.0	3018	2.4178	15.6912	16.6396
2.4839	3.0	4527	2.1905	18.1971	16.4884
2.3044	4.0	6036	2.0272	20.197	16.4735
2.1943	5.0	7545	1.9012	22.2265	16.4266
2.0669	6.0	9054	1.7984	23.7711	16.353
1.985	7.0	10563	1.7100	24.986	16.284
1.9024	8.0	12072	1.6346	26.1758	16.217
1.8484	9.0	13581	1.5692	27.2782	16.1924
1.7761	10.0	15090	1.5111	28.2761	16.144
1.733	11.0	16599	1.4599	29.2184	16.2438
1.6772	12.0	18108	1.4150	30.0026	16.1949
1.6297	13.0	19617	1.3743	30.7839	16.1565
1.5918	14.0	21126	1.3370	31.4921	16.1323
1.5548	15.0	22635	1.3038	32.0621	16.076
1.5333	16.0	24144	1.2743	32.6881	16.0078
1.5145	17.0	25653	1.2478	33.3794	16.1228
1.4826	18.0	27162	1.2240	33.8335	16.0809
1.4488	19.0	28671	1.2021	34.2819	16.0479
1.4386	20.0	30180	1.1829	34.7206	16.0578
1.4127	21.0	31689	1.1660	35.031	16.0717
1.4089	22.0	33198	1.1510	35.4142	16.0391
1.3922	23.0	34707	1.1380	35.6777	16.0461
1.377	24.0	36216	1.1273	35.95	16.0569
1.3598	25.0	37725	1.1175	36.2435	16.0426
1.3515	26.0	39234	1.1097	36.4009	16.0247
1.3441	27.0	40743	1.1042	36.4815	16.0447
1.3412	28.0	42252	1.1001	36.6092	16.0489
1.3527	29.0	43761	1.0976	36.6703	16.0383
1.3397	30.0	45270	1.0968	36.687	16.039

Framework versions

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