|
|
--- |
|
|
datasets: |
|
|
- ccdv/arxiv-summarization |
|
|
metrics: |
|
|
- rouge |
|
|
--- |
|
|
## distilbart-sum-arxiv |
|
|
This model is a fine-tuned version of [sshleifer/distilbart-xsum-12-6](https://huggingface.co/sshleifer/distilbart-cnn-12-6) on a subset of the [ccdv/arxiv-summarization dataset](https://huggingface.co/datasets/ccdv/arxiv-summarization). It achieves the following results on the evaluation set: |
|
|
* Loss: 2.420 |
|
|
* Rouge1: 42.185 |
|
|
* Rouge2: 15.481 |
|
|
* RougeL: 24.440 |
|
|
* RougeLSum: 24.260 |
|
|
|
|
|
## Model description |
|
|
This model is a distilled version of BART with 306M parameters (vs. 406 for the BART model), but it is 1.68 times faster than BART at inference. It has been trained on 60_000 samples and has a limitation of 1024 tokens. |
|
|
|
|
|
## Intended uses & limitations |
|
|
Since this model has been trained on scientific papers, it may perform poorly when attempting to summarize other types of content. |
|
|
|
|
|
```python |
|
|
from transformers import pipeline |
|
|
summarizer = pipeline("text2text-generation", model="AdamCodd/distilbart-sum-arxiv") |
|
|
paper = "Scientific paper..." |
|
|
result = summarizer(paper) |
|
|
print(result) |
|
|
``` |
|
|
|
|
|
## Training and evaluation data |
|
|
|
|
|
More information needed |
|
|
|
|
|
## Training procedure |
|
|
|
|
|
### Training hyperparameters |
|
|
|
|
|
The following hyperparameters were used during training: |
|
|
- learning_rate: 3e-05 |
|
|
- train_batch_size: 4 |
|
|
- eval_batch_size: 4 |
|
|
- seed: 1270 |
|
|
- optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-08 |
|
|
- lr_scheduler_type: linear |
|
|
- lr_scheduler_warmup_steps: 150 |
|
|
- num_epochs: 1 |
|
|
|
|
|
### Training results |
|
|
|
|
|
| key | value | |
|
|
| --- | ----- | |
|
|
| eval_rouge1 | 42.185638427734375 | |
|
|
| eval_rouge2 | 15.481599807739258 | |
|
|
| eval_rougeL | 24.440900802612305 | |
|
|
| eval_rougeLsum | 24.260608673095703 | |
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- Transformers 4.33.0 |
|
|
- Pytorch lightning 2.0.8 |
|
|
- Tokenizers 0.13.3 |
