UER's picture uer

Chinese RoBERTa Miniatures

Model description

This is the set of 24 Chinese RoBERTa models pre-trained by UER-py.

Turc et al. have shown that the standard BERT recipe is effective on a wide range of model sizes. Following their paper, we released the 24 Chinese RoBERTa models. In order to facilitate users to reproduce the results, we used the publicly available corpus and provided all training details.

You can download the 24 Chinese RoBERTa miniatures either from the UER-py Github page, or via HuggingFace from the links below:

Here are scores on the devlopment set of six Chinese tasks:

Model Score douban chnsenticorp lcqmc tnews(CLUE) iflytek(CLUE) ocnli(CLUE)
RoBERTa-Tiny 72.3 83.0 91.4 81.8 62.0 55.0 60.3
RoBERTa-Mini 75.7 84.8 93.7 86.1 63.9 58.3 67.4
RoBERTa-Small 76.8 86.5 93.4 86.5 65.1 59.4 69.7
RoBERTa-Medium 77.8 87.6 94.8 88.1 65.6 59.5 71.2
RoBERTa-Base 79.5 89.1 95.2 89.2 67.0 60.9 75.5

For each task, we selected the best fine-tuning hyperparameters from the lists below, and trained with the sequence length of 128:

  • epochs: 3, 5, 8
  • batch sizes: 32, 64
  • learning rates: 3e-5, 1e-4, 3e-4

How to use

You can use this model directly with a pipeline for masked language modeling (take the case of RoBERTa-Medium):

>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/chinese_roberta_L-8_H-512')
>>> unmasker("中国的首都是[MASK]京。")
    {'sequence': '[CLS] 中 国 的 首 都 是 北 京 。 [SEP]', 
     'score': 0.8701988458633423, 
     'token': 1266, 
     'token_str': '北'},
    {'sequence': '[CLS] 中 国 的 首 都 是 南 京 。 [SEP]',
     'score': 0.1194809079170227, 
     'token': 1298, 
     'token_str': '南'},
    {'sequence': '[CLS] 中 国 的 首 都 是 东 京 。 [SEP]', 
     'score': 0.0037803512532263994, 
     'token': 691, 
     'token_str': '东'},
    {'sequence': '[CLS] 中 国 的 首 都 是 普 京 。 [SEP]',
     'score': 0.0017127094324678183, 
     'token': 3249,
     'token_str': '普'},
    {'sequence': '[CLS] 中 国 的 首 都 是 望 京 。 [SEP]',
     'score': 0.001687526935711503,
     'token': 3307, 
     'token_str': '望'}

Here is how to use this model to get the features of a given text in PyTorch:

from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('uer/chinese_roberta_L-8_H-512')
model = BertModel.from_pretrained("uer/chinese_roberta_L-8_H-512")
text = "用你喜欢的任何文本替换我。"
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

and in TensorFlow:

from transformers import BertTokenizer, TFBertModel
tokenizer = BertTokenizer.from_pretrained('uer/chinese_roberta_L-8_H-512')
model = TFBertModel.from_pretrained("uer/chinese_roberta_L-8_H-512")
text = "用你喜欢的任何文本替换我。"
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)

Training data

CLUECorpusSmall is used as training data. We found that models pre-trained on CLUECorpusSmall outperform those pre-trained on CLUECorpus2020, although CLUECorpus2020 is much larger than CLUECorpusSmall.

Training procedure

Models are pre-trained by UER-py on Tencent Cloud TI-ONE. We pre-train 1,000,000 steps with a sequence length of 128 and then pre-train 250,000 additional steps with a sequence length of 512. We use the same hyper-parameters on different model sizes.

Taking the case of RoBERTa-Medium


python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
                      --vocab_path models/google_zh_vocab.txt \
                      --dataset_path cluecorpussmall_seq128_dataset.pt \
                      --processes_num 32 --seq_length 128 \
                      --dynamic_masking --target mlm
python3 pretrain.py --dataset_path cluecorpussmall_seq128_dataset.pt \
                    --vocab_path models/google_zh_vocab.txt \
                    --config_path models/bert/medium_config.json \
                    --output_model_path models/cluecorpussmall_roberta_medium_seq128_model.bin \
                    --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                    --total_steps 1000000 --save_checkpoint_steps 100000 --report_steps 50000 \
                    --learning_rate 1e-4 --batch_size 64 \
                    --embedding word_pos_seg --encoder transformer --mask fully_visible --target mlm --tie_weights


python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
                      --vocab_path models/google_zh_vocab.txt \
                      --dataset_path cluecorpussmall_seq512_dataset.pt \
                      --processes_num 32 --seq_length 512 \
                      --dynamic_masking --target mlm
python3 pretrain.py --dataset_path cluecorpussmall_seq512_dataset.pt \
                    --pretrained_model_path models/cluecorpussmall_roberta_medium_seq128_model.bin-1000000 \
                    --vocab_path models/google_zh_vocab.txt \
                    --config_path models/bert/medium_config.json \
                    --output_model_path models/cluecorpussmall_roberta_medium_seq512_model.bin \
                    --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                    --total_steps 250000 --save_checkpoint_steps 50000 --report_steps 10000 \
                    --learning_rate 5e-5 --batch_size 16 \
                    --embedding word_pos_seg --encoder transformer --mask fully_visible --target mlm --tie_weights

Finally, we convert the pre-trained model into Huggingface's format:

python3 scripts/convert_bert_from_uer_to_huggingface.py --input_model_path models/cluecorpussmall_roberta_medium_seq512_model.bin-250000 \
                                                        --output_model_path pytorch_model.bin \
                                                        --layers_num 8 --target mlm

BibTeX entry and citation info

  title={Bert: Pre-training of deep bidirectional transformers for language understanding},
  author={Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
  journal={arXiv preprint arXiv:1810.04805},

  title={Roberta: A robustly optimized bert pretraining approach},
  author={Liu, Yinhan and Ott, Myle and Goyal, Naman and Du, Jingfei and Joshi, Mandar and Chen, Danqi and Levy, Omer and Lewis, Mike and Zettlemoyer, Luke and Stoyanov, Veselin},
  journal={arXiv preprint arXiv:1907.11692},

  title={Well-Read Students Learn Better: On the Importance of Pre-training Compact Models},
  author={Turc, Iulia and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
  journal={arXiv preprint arXiv:1908.08962v2 },

  title={UER: An Open-Source Toolkit for Pre-training Models},
  author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
  journal={EMNLP-IJCNLP 2019},
Downloads last month
Hosted inference API
Mask token: [MASK]