bart-samsung-5

This model is a fine-tuned version of facebook/bart-base on the samsum dataset. It achieves the following results on the evaluation set:

• Loss: 1.4959
• Rouge1: 48.4734
• Rouge2: 25.3475
• Rougel: 40.9144
• Rougelsum: 44.7797
• Gen Len: 18.22

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: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
1.6107	1.0	1841	1.5390	47.1407	24.384	40.4826	43.4437	17.5513
1.5528	2.0	3682	1.4971	48.5483	25.1562	41.1806	44.7254	18.3521
1.4225	3.0	5523	1.5013	48.2461	25.2181	40.9022	44.4942	18.0844
1.3266	4.0	7364	1.4976	48.8949	25.4367	41.2355	45.0961	18.2359
1.2635	5.0	9205	1.4959	48.4734	25.3475	40.9144	44.7797	18.22

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.0+cu116
• Datasets 2.7.1
• Tokenizers 0.13.2