mbart-large-50-finetuned-ar-wikilingua

This model is a fine-tuned version of facebook/mbart-large-50 on the wiki_lingua dataset. It achieves the following results on the evaluation set:

• Loss: 4.0001
• Rouge-1: 22.11
• Rouge-2: 7.33
• Rouge-l: 19.75
• Gen Len: 59.4
• Bertscore: 68.9

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: 1e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 250
• num_epochs: 8
• label_smoothing_factor: 0.1

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge-1	Rouge-2	Rouge-l	Gen Len	Bertscore
5.2671	1.0	5111	4.6414	18.37	5.63	16.32	96.39	65.12
4.5375	2.0	10222	4.3144	20.49	6.64	18.35	95.44	65.79
4.308	3.0	15333	4.1592	21.16	7.09	18.85	67.75	67.65
4.1562	4.0	20444	4.0812	21.59	7.31	19.42	68.66	68.02
4.0749	5.0	25555	4.0409	21.99	7.42	19.82	66.4	68.05
4.0271	6.0	30666	4.0183	22.04	7.42	19.64	56.88	68.95
3.9991	7.0	35777	4.0042	22.05	7.35	19.71	55.75	68.94
3.9833	8.0	40888	4.0001	22.12	7.39	19.78	55.72	69.0

Framework versions

• Transformers 4.18.0
• Pytorch 1.10.0+cu111
• Datasets 2.1.0
• Tokenizers 0.12.1