metadata

base_model: google/pegasus-cnn_dailymail
model-index:
  - name: pegasus-samsum
    results: []
datasets:
  - Samsung/samsum
language:
  - en
metrics:
  - rouge
pipeline_tag: summarization
library_name: transformers

pegasus-samsum

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

Loss: 1.3839

Intended uses & limitations

Intended uses:

Dialogue summarization (e.g., chat logs, meetings)
Text summarization for conversational datasets

Limitations:

May struggle with very long conversations or non-dialogue text.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
gradient_accumulation_steps: 16
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.6026	0.5431	500	1.4875
1.4737	1.0861	1000	1.4040
1.4735	1.6292	1500	1.3839

Test results

rouge1	rouge2	rougeL	rougeLsum Loss
0.427614	0.200571	0.340648	0.340738

How to use

You can use this model with the transformers library for dialogue summarization. Here's an example in Python:

from transformers import pipeline
import torch

device = 0 if torch.cuda.is_available() else -1
pipe = pipeline("summarization",
                model="seddiktrk/pegasus-samsum",
                device=device)

custom_dialogue = """\
Seddik: Hey, have you tried using PEGASUS for summarization?
John: Yeah, I just started experimenting with it last week!
Seddik: It's pretty powerful, especially for abstractive summaries.
John: I agree! The results are really impressive.
Seddik: I was thinking of using it for my next project. Want to collaborate?
John: Absolutely! We could make some awesome improvements together.
Seddik: Perfect, let's brainstorm ideas this weekend.
John: Sounds like a plan!
"""

# Summarize dialogue
gen_kwargs = {"length_penalty": 0.8, "num_beams":8, "max_length": 128}
print(pipe(custom_dialogue, **gen_kwargs)[0]["summary_text"])

Example Output

John started using PEG for summarization last week. Seddik is thinking of using it for his next project.
John and Seddik will brainstorm ideas this weekend.

Framework versions

Transformers 4.44.0
Pytorch 2.4.0
Datasets 2.21.0
Tokenizers 0.19.1