mt5-base-finetuned-fa

This model is a fine-tuned version of google/mt5-base on the pn_summary dataset. It achieves the following results on the evaluation set:

• Loss: 2.6477
• Rouge-1: 33.7
• Rouge-2: 21.28
• Rouge-l: 31.69
• Gen Len: 19.0
• Bertscore: 74.52

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: 0.0005
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 250
• num_epochs: 5
• label_smoothing_factor: 0.1

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge-1	Rouge-2	Rouge-l	Gen Len	Bertscore
3.3828	1.0	1875	2.8114	32.17	19.47	30.12	18.99	74.25
2.8204	2.0	3750	2.7080	32.67	19.92	30.56	19.0	74.31
2.6907	3.0	5625	2.6724	33.22	20.44	31.11	19.0	74.47
2.6029	4.0	7500	2.6513	33.46	20.75	31.38	19.0	74.54
2.5414	5.0	9375	2.6477	33.68	20.91	31.62	19.0	74.58

Framework versions

• Transformers 4.19.4
• Pytorch 1.11.0+cu113
• Datasets 2.2.2
• Tokenizers 0.12.1