xlmroberta2xlmroberta-finetune-summarization-ar

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

• Loss: 4.1298
• Rouge-1: 21.69
• Rouge-2: 8.73
• Rouge-l: 19.52
• Gen Len: 19.96
• Bertscore: 71.0

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: 5e-05
• 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: 10
• label_smoothing_factor: 0.1

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge-1	Rouge-2	Rouge-l	Gen Len	Bertscore
8.0645	1.0	1172	7.3567	8.22	0.66	7.94	20.0	58.18
7.2042	2.0	2344	6.6058	12.12	2.19	11.4	20.0	63.24
6.4168	3.0	3516	5.8784	16.46	4.31	15.15	20.0	66.3
5.4622	4.0	4688	4.7931	17.6	5.87	15.9	19.99	69.21
4.7829	5.0	5860	4.4418	19.17	6.74	17.22	19.98	70.23
4.4829	6.0	7032	4.2950	19.8	7.11	17.74	19.98	70.38
4.304	7.0	8204	4.2155	20.71	7.59	18.56	19.98	70.66
4.1778	8.0	9376	4.1632	21.1	7.94	18.99	19.98	70.86
4.0886	9.0	10548	4.1346	21.44	8.03	19.28	19.98	70.93
4.0294	10.0	11720	4.1298	21.51	8.14	19.33	19.98	71.02

Framework versions

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