BARTO (base-sized model) for en-es translation

This model is a fine-tuned version of BARTO on a small portion of WMT13 es-en dataset. It achieves the following results on the evaluation set:

• Loss: 1.7356
• Bleu: 30.2194
• Gen Len: 30.2714

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.005
• train_batch_size: 96
• eval_batch_size: 96
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 384
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 40000
• training_steps: 5000

Framework versions

• Transformers 4.33.0.dev0
• Pytorch 2.0.1
• Datasets 2.14.4
• Tokenizers 0.13.3