mt5-base-finetune-ar-xlsum

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

• Loss: 3.2546
• Rouge-1: 22.2
• Rouge-2: 9.57
• Rouge-l: 20.26
• Gen Len: 19.0
• Bertscore: 71.43

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: 16
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 250
• num_epochs: 10
• label_smoothing_factor: 0.1

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge-1	Rouge-2	Rouge-l	Gen Len	Bertscore
4.9261	1.0	585	3.6314	18.19	6.49	16.37	19.0	70.17
3.8429	2.0	1170	3.4253	19.45	7.58	17.73	19.0	70.35
3.6311	3.0	1755	3.3569	20.83	8.54	18.9	19.0	70.89
3.4917	4.0	2340	3.3101	20.77	8.53	18.89	19.0	70.98
3.3873	5.0	2925	3.2867	21.47	9.0	19.54	19.0	71.23
3.3037	6.0	3510	3.2693	21.41	9.0	19.5	19.0	71.21
3.2357	7.0	4095	3.2581	22.05	9.36	20.04	19.0	71.43
3.1798	8.0	4680	3.2522	22.21	9.56	20.23	19.0	71.41
3.1359	9.0	5265	3.2546	22.27	9.58	20.23	19.0	71.46
3.0997	10.0	5850	3.2546	22.2	9.57	20.26	19.0	71.43

Framework versions

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