File size: 8,455 Bytes
b0c1531 15b9d4a b0c1531 15b9d4a b0c1531 15b9d4a b0c1531 15b9d4a b0c1531 15b9d4a |
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 |
Model card for RoGPT2-medium
---
language:
- ro
---
# RoGPT2: Romanian GPT2 for text generation
All models are available:
* [RoBERT-base](https://huggingface.co/readerbench/RoGPT2-base)
* [RoBERT-medium](https://huggingface.co/readerbench/RoGPT2-medium)
* [RoBERT-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, TFAutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained('readerbench/RoGPT2-medium')
model = TFAutoModelWithLMHead.from_pretrained('readerbench/RoGPT2-medium')
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, AutoModelWithLMHead
tokenizer = AutoTokenizer.from_pretrained('readerbench/RoGPT2-medium')
model = AutoModelWithLMHead.from_pretrained('readerbench/RoGPT2-medium')
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 | Block 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 | - | 72.7 | 59.9 |
|XLM-R Large | - |**83.6**|**69.7**|
|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
---
Niculescu, M. A., Ruseti, S., and Dascalu, M. (submitted). RoGPT2: Romanian GPT2 for Text Generation
|