pegasus-base-arxiv-TitleGeneration

This model is a fine-tuned version of google/pegasus-xsum on the arxiv dataset. It achieves the following results on the evaluation set:

• Loss: 2.8170
• Rouge1: 41.7224
• Rouge2: 22.4944
• Rougel: 38.154
• Rougelsum: 38.1733
• Gen Len: 10.976

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: 2
• eval_batch_size: 2
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
4.106	0.2	500	3.4397	33.3811	15.877	30.4348	30.4856	11.167
3.9194	0.4	1000	3.3273	36.1775	18.1453	33.0183	33.0809	10.251
3.5897	0.6	1500	3.1088	37.555	18.5533	34.512	34.575	10.514
3.4344	0.8	2000	2.9730	39.1491	20.1873	35.4581	35.5301	11.307
3.1704	1.0	2500	2.8170	41.7224	22.4944	38.154	38.1733	10.976

Framework versions

• Transformers 4.39.3
• Pytorch 2.1.2
• Datasets 2.18.0
• Tokenizers 0.15.2