Edit model card

AraBART-finetuned-ar

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

  • Loss: 3.7449
  • Rouge-1: 31.08
  • Rouge-2: 14.68
  • Rouge-l: 27.36
  • Gen Len: 19.64
  • Bertscore: 73.86

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: 16
  • eval_batch_size: 16
  • 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: 10
  • label_smoothing_factor: 0.1

Training results

Training Loss Epoch Step Validation Loss Rouge-1 Rouge-2 Rouge-l Gen Len Bertscore
4.4318 1.0 2345 3.7996 28.93 13.2 25.56 19.51 73.17
4.0338 2.0 4690 3.7483 30.29 14.24 26.73 19.5 73.59
3.8586 3.0 7035 3.7281 30.44 14.44 26.92 19.75 73.58
3.7289 4.0 9380 3.7204 30.55 14.49 26.88 19.66 73.73
3.6245 5.0 11725 3.7199 30.73 14.63 27.11 19.69 73.68
3.5392 6.0 14070 3.7221 30.85 14.65 27.21 19.7 73.77
3.4694 7.0 16415 3.7286 31.08 14.8 27.41 19.62 73.84
3.4126 8.0 18760 3.7384 31.06 14.77 27.41 19.64 73.82
3.3718 9.0 21105 3.7398 31.18 14.89 27.49 19.67 73.87
3.3428 10.0 23450 3.7449 31.19 14.88 27.44 19.68 73.87

Framework versions

  • Transformers 4.18.0
  • Pytorch 1.10.0+cu111
  • Datasets 2.1.0
  • Tokenizers 0.12.1
Downloads last month
73
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.