File size: 911 Bytes
74d757e
 
 
 
890b687
74d757e
8f8153d
be7c2d8
8f8153d
 
 
 
 
c120b2a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
---
tags:
- summarization
license: mit
thumbnail: https://en.wikipedia.org/wiki/Bart_Simpson#/media/File:Bart_Simpson_200px.png
---
# BART for Gigaword
 - This model was created by fine-tuning the `facebook/bart-large-cnn` weights (also on HuggingFace) for the Gigaword dataset. The model was fine-tuned on the Gigaword training set for 3 epochs, and the model with the highest ROUGE-1 score on the training set batches was kept.
 - The BART Tokenizer for CNN-Dailymail was used in the fine-tuning process and that is the tokenizer that will be loaded automatically when doing:
```
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("a1noack/bart-large-gigaword")
```
 - This model achieves ROUGE-1 / ROUGE-2 / ROUGE-L of 37.28 / 18.58 / 34.53 on the Gigaword test set; this is pretty good when compared to PEGASUS, `google/pegasus-gigaword`, which achieves 39.12 / 19.86 / 36.24.