Hugging Face's logo

readerbench
/
RoGPT2-large

Text Generation
Transformers
PyTorch
TensorFlow
gpt2
Inference Endpoints
text-generation-inference
Model card Files Community
1
RoGPT2-large
File size: 8,693 Bytes 
bb89bc3
 
 
7100155
bb89bc3
 
 
 
 
 
bd22d3e
 
 
bb89bc3
7100155
 
bb89bc3
 
7100155
 
0a955f1
bb89bc3
 
0a955f1
bb89bc3
 
 
7100155
 
0a955f1
7100155
 
e60b3cd
7100155
 
 
bb89bc3
7100155
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d3292f
7100155
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5d3292f
 
7100155
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
839061d
bd22d3e
 
 
 
 
 
 
839061d
 
 
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
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
 
Model card for RoGPT2-large

---
language:
- ro
---

# RoGPT2: Romanian GPT2 for text generation

All models are available:
* [RoGPT2-base](https://huggingface.co/readerbench/RoGPT2-base)
* [RoGPT2-medium](https://huggingface.co/readerbench/RoGPT2-medium)
* [RoGPT2-large](https://huggingface.co/readerbench/RoGPT2-large)

For code and evaluation check out [GitHub](https://github.com/readerbench/RoGPT2).

#### How to use

```python
# TensorFlow
from transformers import AutoTokenizer, TFAutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained('readerbench/RoGPT2-large')
model = TFAutoModelForCausalLM.from_pretrained('readerbench/RoGPT2-large')
inputs = tokenizer.encode("Este o zi de vara", return_tensors='tf')
text = model.generate(inputs, max_length=1024,  no_repeat_ngram_size=2)
print(tokenizer.decode(text[0]))

# PyTorch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained('readerbench/RoGPT2-large')
model = AutoModelForCausalLM.from_pretrained('readerbench/RoGPT2-large')
inputs = tokenizer.encode("Este o zi de vara", return_tensors='pt')
text = model.generate(inputs, max_length=1024,  no_repeat_ngram_size=2)
print(tokenizer.decode(text[0]))
```

## Training

---

### Corpus Statistics

| Corpus | Total size | Number of words | Number of sentences |
|:------:|:----------:|:---------------:|:-------------------:|
|OSCAR| 11.54 GB | 1745M | 48.46M |
|Wiki-Ro | 0.46 GB | 68M | 1.79M |
|Debates | 0.5 GB | 73M | 3.61M |
|Books | 4.37 GB | 667M | 37.39M |
|News | 0.15 GB | 23M | 0.77M |

### Training Statistics

| Version | Number of parameters | Number of epoch | Duration of an epoch | Context size | Batch size | PPL |
|:-------:|:--------------------:|:---------------:|:--------------------:|:----------:|:----------:|:---:|
| Base   | 124M | 15 | 7h      | 1024 | 72 | 22.96    |
| Medium | 354M | 10 | 22h     | 1024 | 24 | 17.64    |
| Large  | 774M | 5  | **45h** | 512  | 16 | **16.77**|

## Evaluation

---

### 1. MOROCO

|      Model        | Dialect | Md to Ro | Ro to Md |
|:-----------------:|:-------:|:--------:|:--------:|
| KRR + SK          | 94.06   | 67.59    | 75.47    |
| BERT-base-ro      | 95.98   | 69.90    | 78.08    |
| RoBERT-small      | 95.76   | 69.05    | 80.15    |
| RoBERT-base       |**97.24**| 68.80    | 82.37    |
| RoBERT-large      | 97.21   | 69.50    | **83.26**|
| RoGPT2-base       | 96.69   | 69.82    | 77.55    |
| RoGPT2-medium     | 96.42   | 69.77    | 80.51    |
| RoGPT2-large      | 96.93   |**71.07** | 82.56    |

### 2. LaRoSeDa

| Model        | Binary: Accuracy | Binary: F1-Score | Multi-Class: Accuracy | Multi-Class: F1-Score |
|:------------:|:----------------:|:----------------:|:---------------------:|:---------------------:|
|BERT-base-ro  | 98.07            | 97.94            | -                     |79.61                  |
| RoDiBERT     |**98.40**         |**98.31**         | -                     |83.01                  |
| RoBERT-small | 97.44            | 97.43            | 89.30                 |84.23                  |
| RoBERT-base  | 98.27            | 98.26            | 90.59                 |86.27                  |
| RoBERT-large | 98.20            | 98.19            |**90.93**              |**86.63**              |
| RoGPT2-base  | 97.89            | 97.88            |89.65                  |84.68                  |
|RoGPT2-medium | 98.03            |98.04             | 90.29                 | 85.37                 |
| RoGPT2-large | 98.06            |98.07             | 90.26                 | 84.89                 |

### 3. RoSTS

| Model        | Spearman dev-set | Spearman test-set | Pearson dev-set | Pearson test-set |
|:------------:|:----------------:|:-----------------:|:---------------:|:----------------:|
|BERT-base-ro  | 84.26            | 80.86             | 84.59           | 81.59            |
|RoDiBERT      | 77.07            | 71.47             | 77.13           | 72.25            |
|RoBERT-small  | 82.06            | 78.06             | 81.66           | 78.49            |
|RoBERT-base   | 84.93            | 80.39             | 85.03           | 80.39            |
|RoBERT-large  |**86.25**         |**83.15**          |**86.58**        |**83.76**         |
|RoGPT2-base   | 83.51            | 79.77             | 83.74           | 80.56            |
|RoGPT2-medium | 85.75            | 82.25             | 86.04           | 83.16            |
|RoGPT2-large  | 85.70            | 82.64             | 86.14           | 83.46            |

### 4. WMT16

| Model        | Decoder method | Ro-En  | En-Ro  |
|:------------:|:--------------:|:------:|:------:|
|mBART         | -              |**38.5**|**38.5**|
|OpenNMT       | -              |  -     | 24.7   |
|RoGPT2-base   |Greedy          | 30.37  | 20.27  |
|RoGPT2-base   |Beam-search-4   | 31.26  | 22.31  |
|RoGPT2-base   |Beam-search-8   | 31.39  | 22.95  |
|RoGPT2-medium |Greedy          | 32.48  | 22.18  |
|RoGPT2-medium |Beam-search-4   | 34.08  | 24.03  |
|RoGPT2-medium |Beam-search-8   | 34.16  | 24.13  |
|RoGPT2-large  |Greedy          | 33.69  | 23.31  |
|RoGPT2-large  |Beam-search-4   |34.40   |24.23   |
|RoGPT2-large  |Beam-search-8   |34.51   |24.32   | 

### 5. XQuAD
| Model        |Decoder method |  EM   | F1-Score |
|:------------:|:-------------:|:-----:|:--------:|
|BERT-base-ro  | -             | 47.89 | 63.74    |
|RoDiBERT      | -             | 21.76 | 34.57    |
|RoBERT-small  | -             | 30.84 | 45.17    |
|RoBERT-base   | -             | 53.52 | 70.04    |
|RoBERT-large  | -             | 55.46 |  69.64   |
|mBERT         | -             | 59.9  |  72.7  |
|XLM-R Large   | -             |**69.7**|**83.6**|
|RoGPT2-base   |  Greedy       | 23.69 | 35.97    |
|RoGPT2-base   | Beam-search-4 | 24.11 | 35.27    |
|RoGPT2-medium | Greedy        | 29.66 | 44.74    |
|RoGPT2-medium | Beam-search-4 | 31.59 | 45.32 |
|RoGPT2-large | Greedy | 29.74 | 42.98 |
|RoGPT2-large | Beam-search-4 | 29.66 | 43.05 |
|RoGPT2-base-en-ro | Greedy | 23.86 | 34.27 |
|RoGPT2-base-en-ro | Beam-search-4 | 25.04 | 34.51 |
|RoGPT2-medium-en-ro | Greedy | 27.05 | 39.75 |
|RoGPT2-medium-en-ro | Beam-search-4 | 27.64 | 39.11 |
|RoGPT2-large-en-ro | Greedy | 28.40 | 39.79 |
|RoGPT2-large-en-ro | Beam-search-4 | 28.73 | 39.71 |
|RoGPT2-large-en-ro-mask | Greedy | 31.34 | 44.71 |
|RoGPT2-large-en-ro-mask|  Beam-search-4 | 31.59 | 43.53 |

### 6. Wiki-Ro: LM

| Model        | PPL dev | PPL test |
|:------------:|:-------:|:--------:|
|BERT-base-ro | 29.0897 | 28.0043|
|RoGPT2-base | 34.3795 | 33.7460|
|RoGPT2-medium | 23.7879 | 23.4581|
|RoGPT2-large | **21.7491** | **21.5200** |

### 7. RoGEC

| Model | Decoder mothod |  P  |  R  |  F<sub>0.5</sub>   |
|:-----:|:--------------:|:---:|:---:|:------:|
|Transformer-tiny | Beam-search | 53.53 | 26.36 | 44.38 |
|Transformer-base Finetuning | Beam-search | 56.05 | 46.19 | 53.76 |
|Transformer-base Finetuning | Beam-search-LM | 50.68 | 45.39 | 49.52 |
|Transformer-base Finetuning | Beam-search-norm-LM | 51.06 | 45.43 | 49.83 |
|RoGPT2-base | Greedy | 59.02 | 49.35 | 56.80 |
|RoGPT2-base | Beam-search-4 | 65.23 | 49.26 | 61.26 |
|RoGPT2-base  |Beam-search-8 | 65.88 | 49.64 | 61.84 |
|RoGPT2-medium | Greedy | 69.97 | 57.94 | 67.18 |
|RoGPT2-medium | Beam-search-4 | **72.46** | **57.99** | **69.01** |
|RoGPT2-medium | Beam-search-8 | 72.24 | 57.69 | 68.77 |
|RoGP2-large | Greedy | 61.90 | 49.09 | 58.83 |
|RoGP2-large | Beam-search-4 | 65.24 | 49.43 | 61.32 |
|RoGP2-large | Beam-search-8 | 64.96 | 49.22 | 61.06 |
|RoGPT2-base* | Greedy | 68.67 | 49.60 | 63.77 |
|RoGPT2-base* | Beam-search-4 | 71.16 | 50.53 | 65.79 |
|RoGPT2-base* | Beam-search-8 | 71.68 | 50.65 | 66.18 |
|RoGPT2-medium* | Greedy | 58.21 | 43.32 | 54.47 |
|RoGPT2-medium* | Beam-search-4 | 68.31 | 43.78 | 61.43 |
|RoGPT2-medium* | Beam-search-8 | 68.68 | 43.99 | 61.75 |
|RoGPT2-large* | Greedy | 64.86 | 41.30 | 58.22 |
|RoGPT2-large* | Beam-search-4 | 65.57 | 41.00 | 58.55 |
|RoGPT2-large* | Beam-search-8 | 65.44 | 41.09 | 58.50 |

**__Note__**: * the models were trained using the dataset of 3,000,000 artificially generated pairs

## Acknowledgments

---

Research supported with [Cloud TPUs](https://cloud.google.com/tpu/) from Google's [TensorFlow Research Cloud (TFRC)](https://www.tensorflow.org/tfrc)


## How to cite

---
```bibtex
@inproceedings{niculescu2021rogpt2,
  title={RoGPT2: Romanian GPT2 for Text Generation},
  author={Niculescu, Mihai Alexandru and Ruseti, Stefan and Dascalu, Mihai},
  booktitle={2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI)},
  pages={1154--1161},
  year={2021},
  organization={IEEE}
}
```