File size: 4,270 Bytes
ece032c
 
 
 
 
57c555c
ece032c
 
591097e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
57c555c
591097e
 
 
57c555c
 
 
 
 
 
 
ece032c
 
 
 
 
 
 
 
591097e
ece032c
7427deb
 
 
 
 
 
 
 
 
 
 
 
ece032c
 
 
7427deb
 
 
a7e8c2c
 
e685f90
dadcbe0
591097e
e685f90
 
 
 
 
 
 
 
 
 
08748c3
e685f90
 
 
 
 
6558c20
555d47d
 
 
333f437
510a7fc
1a89701
 
 
 
 
 
 
555d47d
 
591097e
ece032c
 
57c555c
 
 
ece032c
 
 
333f437
 
 
 
 
 
 
 
 
 
 
 
 
ece032c
 
c0ec00c
 
 
 
 
 
 
 
 
 
 
 
 
 
ece032c
 
 
 
 
57c555c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
---
base_model: google/pegasus-x-base
tags:
- generated_from_trainer
datasets:
- ccdv/arxiv-summarization
model-index:
- name: Paper-Summarization-ArXiv
  results:
  - task:
      name: Summarization
      type: summarization
    dataset:
      name: ccdv/arxiv-summarization
      type: ccdv/arxiv-summarization
      config: section
      split: test
      args: section
    metrics:
    - name: ROUGE-1
      type: rouge
      value: 43.2305
    - name: ROUGE-2
      type: rouge
      value: 16.6571
    - name: ROUGE-L
      type: rouge
      value: 24.4315
    - name: ROUGE-LSum
      type: rouge
      value: 33.9399
license: bigscience-openrail-m
language:
- en
metrics:
- rouge
library_name: transformers
pipeline_tag: summarization
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# Paper-Summarization-ArXiv

This model is a fine-tuned version of [google/pegasus-x-base](https://huggingface.co/google/pegasus-x-base) on the arxiv-summarization dataset.


**Base Model**: [Pegasus-x-base (State-of-the-art for Long Context Summarization)](https://huggingface.co/google/pegasus-x-base)

**Finetuning Dataset**: 
- We used **full of ArXiv Dataset (Cohan et al., 2018, NAACL-HLT 2018)** [[PDF]](https://arxiv.org/abs/1804.05685)
  - (Full length is 200,000+)

**GPU**: (RTX A6000) x 1

**Train time**: About 120 hours for 5 epochs

**Test time**: About 8 hours for test dataset.


## Intended uses & limitations

- **Research Paper Summarization**



## Compare to Baseline
- Pegasus-X-base **zero-shot** Performance:
  - R-1 | R-2 | R-L | R-LSUM : 6.2269 | 0.7894 | 4.6905 | 5.4591

- **This model**


  - R-1 | R-2 | R-L | R-LSUM : 43.2305 | 16.6571 | 24.4315 | 33.9399 at 
  ```(python)
  model.generate(input_ids =inputs["input_ids"].to(device),
                              attention_mask=inputs["attention_mask"].to(device),
                              length_penalty=1, num_beams=2, max_length=128*4,min_length=150, no_repeat_ngram_size= 3, top_k=25,top_p=0.95)
    
  ```
  - R-1 | R-2 | R-L | R-LSUM : 40.8486 | 16.3717 | 25.2937 | 33.6923 (refer to **PEGASUS-X's [paper](https://arxiv.org/pdf/2208.04347.pdf)**) at 
  ```(python)
  model.generate(input_ids =inputs["input_ids"].to(device),
                              attention_mask=inputs["attention_mask"].to(device),
                              length_penalty=1, num_beams=1, max_length=128*2,top_p=1)
  ```
  - R-1 | R-2 | R-L | R-LSUM : 38.1317 | 15.0357 | 23.0286 | 30.9938 (**Diverse Beam-Search Decoding**) at
  ```(python)
  model.generate(input_ids =inputs["input_ids"].to(device),
                              attention_mask=inputs["attention_mask"].to(device),
                              num_beam_groups=5,diversity_penalty=1.0,num_beams=5,min_length=150,max_length=128*4)
  ```
  - R-1 | R-2 | R-L | R-LSUM : 43.3017 | 16.6023 | 24.1867 | 33.7019  at
  ```(python)
  model.generate(input_ids =inputs["input_ids"].to(device),
                              attention_mask=inputs["attention_mask"].to(device),
                              length_penalty=1.2, num_beams=4, max_length=128*4,min_length=150, no_repeat_ngram_size= 3, temperature=0.9,top_k=50,top_p=0.92)
   
  ```



## Training procedure

We use huggingface-based environment such as datasets, trainer, etc.


### Training hyperparameters

The following hyperparameters were used during training:

```(python)
learning_rate: 1e-05,
train_batch_size: 1,
eval_batch_size: 1,
seed: 42,
gradient_accumulation_steps: 64,
total_train_batch_size: 64,
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08,
lr_scheduler_type: linear,
lr_scheduler_warmup_steps: 1586,
num_epochs: 5
```


### Training results

| Training Loss | Epoch | Step  | Validation Loss |
|:-------------:|:-----:|:-----:|:---------------:|
| 2.6153        | 1.0   | 3172  | 2.1045          |
| 2.202         | 2.0   | 6344  | 2.0511          |
| 2.1547        | 3.0   | 9516  | 2.0282          |
| 2.132         | 4.0   | 12688 | 2.0164          |
| 2.1222        | 5.0   | 15860 | 2.0127          |





### Framework versions

- Transformers 4.32.1
- Pytorch 2.0.1
- Datasets 2.12.0
- Tokenizers 0.13.2