MarianMix_en-10

This model is a fine-tuned version of Helsinki-NLP/opus-tatoeba-en-ja on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.0752
• Bleu: 14.601
• Gen Len: 45.8087

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: 0.0001
• train_batch_size: 32
• eval_batch_size: 32
• seed: 99
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 10
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Bleu	Gen Len
2.1136	0.44	500	2.0044	0.2655	109.0201
1.1422	0.89	1000	1.7516	1.4123	71.0
0.9666	1.33	1500	1.5219	3.6611	64.6888
0.8725	1.78	2000	1.3606	4.6539	77.1641
0.7655	2.22	2500	1.2586	8.3456	60.3837
0.7149	2.67	3000	1.1953	11.2247	50.5921
0.6719	3.11	3500	1.1541	10.4303	54.3776
0.6265	3.56	4000	1.1186	13.3231	48.283
0.6157	4.0	4500	1.0929	13.8467	46.569
0.5736	4.44	5000	1.0848	14.2731	45.5035
0.5683	4.89	5500	1.0752	14.601	45.8087

Framework versions

• Transformers 4.12.5
• Pytorch 1.9.1
• Datasets 1.17.0
• Tokenizers 0.10.3