distilbart-cnn-6-6-reddit
This model is a fine-tuned version of sshleifer/distilbart-cnn-6-6 on the reddit dataset. It achieves the following results on the evaluation set:
- Loss: 2.9883
- Rouge1: 0.1849
- Rouge2: 0.0437
- Rougel: 0.1273
- Rougelsum: 0.1601
More information and training script
You can find more information about how this model was trained, including the actual training script in this github repository.
Training and evaluation data
I made a split in a train and test set. The test size is 1% of the total dataset, which comes down to about 38k samples.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- 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
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss | Rouge1 | Rouge2 | Rougel | Rougelsum |
|---|---|---|---|---|---|---|---|
| 3.13 | 1.0 | 238116 | 3.2736 | 0.1773 | 0.0392 | 0.1223 | 0.1539 |
| 2.8586 | 2.0 | 476232 | 3.0449 | 0.1846 | 0.0431 | 0.127 | 0.1601 |
| 2.7844 | 3.0 | 714348 | 2.9883 | 0.1849 | 0.0437 | 0.1273 | 0.1601 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.13.1+cu116
- Datasets 2.8.0
- Tokenizers 0.13.2
