Chinese Xlarge Whole Word Masking RoBERTa Model

Model description

This is an xlarge Chinese Whole Word Masking RoBERTa model pre-trained by TencentPretrain introduced in this paper, which inherits UER-py to support models with parameters above one billion, and extends it to a multimodal pre-training framework.

In order to facilitate users in reproducing the results, we used a publicly available corpus and word segmentation tool, and provided all training details.

You can download the model either from the UER-py Modelzoo page, or via HuggingFace from the link roberta-xlarge-wwm-chinese-cluecorpussmall:

How to use

You can use this model directly with a pipeline for masked language modeling:

>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-xlarge-wwm-chinese-cluecorpussmall')
>>> unmasker("北京是[MASK]国的首都。")
[
  {'score': 0.9298505783081055,
   'token': 704,
   'token_str': '中',
   'sequence': '北 京 是 中 国 的 首 都 。'},
  {'score': 0.05041525512933731,
   'token': 2769,
   'token_str': '我',
   'sequence': '北 京 是 我 国 的 首 都 。'},
  {'score': 0.004921116400510073,
   'token': 4862,
   'token_str': '祖',
   'sequence': '北 京 是 祖 国 的 首 都 。'},
  {'score': 0.0020684923510998487,
   'token': 3696,
   'token_str': '民',
   'sequence': '北 京 是 民 国 的 首 都 。'},
  {'score': 0.0018144999630749226,
   'token': 3926,
   'token_str': '清',
   'sequence': '北 京 是 清 国 的 首 都 。'}
]

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/roberta-xlarge-wwm-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-xlarge-wwm-chinese-cluecorpussmall")
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/roberta-xlarge-wwm-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-xlarge-wwm-chinese-cluecorpussmall")
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.

Training procedure

Models are pre-trained by TencentPretrain on Tencent Cloud. We pre-train 500,000 steps with a sequence length of 128 and then pre-train 250,000 additional steps with a sequence length of 512.

jieba is used as word segmentation tool.

Stage1:

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 --data_processor mlm

deepspeed pretrain.py --deepspeed --deepspeed_config models/deepspeed_config.json --dataset_path cluecorpussmall_seq128_dataset.pt \
                      --vocab_path models/google_zh_vocab.txt \
                      --config_path models/bert/xlarge_config.json \
                      --output_model_path models/cluecorpussmall_wwm_roberta_xlarge_seq128_model \
                      --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                      --total_steps 500000 --save_checkpoint_steps 50000 --report_steps 500 \
                      --learning_rate 2e-5 --batch_size 128 --deep_init \
                      --whole_word_masking  --deepspeed_checkpoint_activations \
                      --data_processor mlm --target mlm

Before stage2, we extract fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints:

python3 models/cluecorpussmall_wwm_roberta_xlarge_seq128_model/zero_to_fp32.py models/cluecorpussmall_wwm_roberta_xlarge_seq128_model/ \
                                                                        models/cluecorpussmall_wwm_roberta_xlarge_seq128_model.bin

Stage2:

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 --data_processor mlm

deepspeed pretrain.py --deepspeed --deepspeed_config models/deepspeed_config.json --dataset_path cluecorpussmall_seq512_dataset.pt \
                      --vocab_path models/google_zh_vocab.txt \
                      --config_path models/bert/xlarge_config.json \
                      --pretrained_model_path models/cluecorpussmall_wwm_roberta_xlarge_seq128_model.bin \
                      --output_model_path models/cluecorpussmall_wwm_roberta_xlarge_seq512_model \
                      --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                      --total_steps 250000 --save_checkpoint_steps 50000 --report_steps 500 \
                      --learning_rate 5e-5 --batch_size 32 \
                      --whole_word_masking --deepspeed_checkpoint_activations \
                      --data_processor mlm --target mlm

Then, we extract fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints:

python3 models/cluecorpussmall_wwm_roberta_xlarge_seq512_model/zero_to_fp32.py models/cluecorpussmall_wwm_roberta_xlarge_seq512_model/ \
                                                                               models/cluecorpussmall_wwm_roberta_xlarge_seq512_model.bin

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

python3 scripts/convert_bert_from_tencentpretrain_to_huggingface.py --input_model_path models/cluecorpussmall_wwm_roberta_xlarge_seq512_model.bin \
                                                                    --output_model_path pytorch_model.bin \
                                                                    --layers_num 36 --type mlm

BibTeX entry and citation info

@article{zhao2019uer,
  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},
  pages={241},
  year={2019}
}

@article{zhao2023tencentpretrain,
  title={TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities},
  author={Zhao, Zhe and Li, Yudong and Hou, Cheng and Zhao, Jing and others},
  journal={ACL 2023},
  pages={217},
  year={2023}
}