File size: 8,688 Bytes
4f84de3 febf6a7 4f84de3 a6584cc 4f84de3 febf6a7 4f84de3 febf6a7 4f84de3 febf6a7 cd10edc 4f84de3 cd10edc 4f84de3 febf6a7 cd10edc febf6a7 cd10edc febf6a7 4f84de3 febf6a7 b5e625f febf6a7 9e3c832 febf6a7 0743a18 a6584cc |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 |
Model card for RoGPT2-base
---
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-base')
model = TFAutoModelForCausalLM.from_pretrained('readerbench/RoGPT2-base')
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-base')
model = AutoModelForCausalLM.from_pretrained('readerbench/RoGPT2-base')
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}
}
```
|