modelId
stringlengths 4
112
| lastModified
stringlengths 24
24
| tags
sequence | pipeline_tag
stringclasses 21
values | files
sequence | publishedBy
stringlengths 2
37
| downloads_last_month
int32 0
9.44M
| library
stringclasses 15
values | modelCard
large_stringlengths 0
100k
|
|---|---|---|---|---|---|---|---|---|
typeform/distilbert-base-uncased-mnli | 2021-05-27T09:53:20.000Z | [
"pytorch",
"distilbert",
"text-classification",
"en",
"dataset:multi_nli",
"transformers",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"eval_results_mnli-mm.txt",
"eval_results_mnli.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"train_results.txt",
"trainer_state.json",
"training_args.bin",
"vocab.txt"
] | typeform | 117,594 | transformers | ---
language: en
pipeline_tag: zero-shot-classification
tags:
- distilbert
datasets:
- multi_nli
metrics:
- accuracy
---
# DistilBERT base model (uncased)
This is the [uncased DistilBERT model](https://huggingface.co/distilbert-base-uncased) fine-tuned on [Multi-Genre Natural Language Inference](https://huggingface.co/datasets/multi_nli) (MNLI) dataset for the zero-shot classification task. The model is not case-sensitive, i.e., it does not make a difference between "english" and "English".
## Training
Training is done on a [p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) AWS EC2 instance (1 NVIDIA Tesla V100 GPUs), with the following hyperparameters:
```
$ run_glue.py \
--model_name_or_path distilbert-base-uncased \
--task_name mnli \
--do_train \
--do_eval \
--max_seq_length 128 \
--per_device_train_batch_size 16 \
--learning_rate 2e-5 \
--num_train_epochs 5 \
--output_dir /tmp/distilbert-base-uncased_mnli/
```
## Evaluation results
| Task | MNLI | MNLI-mm |
|:----:|:----:|:----:|
| | 82.0 | 82.0 | |
typeform/distilroberta-base-v2 | 2021-05-20T22:46:35.000Z | [
"pytorch",
"jax",
"roberta",
"masked-lm",
"en",
"dataset:openwebtext",
"transformers",
"license:apache-2.0",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"tokenizer.json"
] | typeform | 19 | transformers | ---
language: en
license: apache-2.0
datasets:
- openwebtext
---
# DistilRoBERTa base model
Forked from https://huggingface.co/distilroberta-base
|
typeform/distilroberta-base | 2021-01-20T14:23:46.000Z | [
"pytorch",
"roberta",
"masked-lm",
"en",
"dataset:openwebtext",
"transformers",
"license:apache-2.0",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"merges.txt",
"pytorch_model.bin",
"tokenizer.json"
] | typeform | 6 | transformers | ---
language: en
license: apache-2.0
datasets:
- openwebtext
---
# DistilRoBERTa base model
Forked from https://huggingface.co/distilroberta-base
|
typeform/mobilebert-uncased-mnli | 2021-02-14T09:11:00.000Z | [
"pytorch",
"mobilebert",
"text-classification",
"en",
"dataset:multi_nli",
"transformers",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.txt"
] | typeform | 5,129 | transformers | ---
language: en
pipeline_tag: zero-shot-classification
tags:
- mobilebert
datasets:
- multi_nli
metrics:
- accuracy
---
# MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices
This model is the Multi-Genre Natural Language Inference (MNLI) fine-turned version of the [uncased MobileBERT model](https://huggingface.co/google/mobilebert-uncased). |
typeform/roberta-large-mnli | 2021-05-20T22:49:31.000Z | [
"tf",
"roberta",
"text-classification",
"multilingual",
"transformers",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"merges.txt",
"modelcard.json",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.json"
] | typeform | 1,393 | transformers | ---
language: multilingual
pipeline_tag: zero-shot-classification
---
# RoBERTa Large Multilanguage |
typeform/squeezebert-mnli | 2021-02-13T19:41:31.000Z | [
"pytorch",
"squeezebert",
"en",
"dataset:mulit_nli",
"transformers",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"tokenizer.json",
"vocab.txt"
] | typeform | 195 | transformers | ---
language: en
pipeline_tag: zero-shot-classification
tags:
- squeezebert
datasets:
- mulit_nli
metrics:
- accuracy
---
# SqueezeBERT |
tyw/test | 2021-05-14T05:13:55.000Z | [] | [
".gitattributes",
"README.md"
] | tyw | 0 | |||
uasoyasser/eefdfgdg | 2021-03-05T15:37:12.000Z | [] | [
".gitattributes",
"README.md"
] | uasoyasser | 0 | https://teacher.desmos.com/activitybuilder/teacherguide/604249659240440d25a27d0c
https://teacher.desmos.com/activitybuilder/teacherguide/604249a365ecd40d30b4ad18
https://teacher.desmos.com/activitybuilder/teacherguide/604249e2cfb0a20d51e13768
https://teacher.desmos.com/activitybuilder/teacherguide/60424a1c9240440d25a27e22
https://teacher.desmos.com/activitybuilder/teacherguide/60424a58cefbd00d5da96390
https://teacher.desmos.com/activitybuilder/teacherguide/60424a90229a7d0cfb807295
https://teacher.desmos.com/activitybuilder/teacherguide/60424ad532e0730c4bdcbbab
https://teacher.desmos.com/activitybuilder/teacherguide/60424b0f1d780b0b7395f36d
https://teacher.desmos.com/activitybuilder/teacherguide/60424c01534b110d262d4d46
https://teacher.desmos.com/activitybuilder/teacherguide/60424c47969a440d13c62ffb
https://teacher.desmos.com/activitybuilder/teacherguide/60424cd7f17f6b0d4550c269
https://teacher.desmos.com/activitybuilder/teacherguide/60424d0dcfb0a20d51e13c97
https://teacher.desmos.com/activitybuilder/teacherguide/60424d5796540a0cf95ff215
https://teacher.desmos.com/activitybuilder/teacherguide/60424d9163a2220bc4c8f2be
https://teacher.desmos.com/activitybuilder/teacherguide/60424e030d98a80d53856ab2
https://teacher.desmos.com/activitybuilder/teacherguide/60424e37ed488c0cfbbaab2f |
||
uclanlp/visualbert-nlvr2-coco-pre | 2021-05-31T11:11:50.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 6 | transformers | ||
uclanlp/visualbert-nlvr2-pre | 2021-05-31T11:12:02.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 10 | transformers | ||
uclanlp/visualbert-nlvr2 | 2021-05-31T11:09:59.000Z | [
"pytorch",
"visual_bert",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 65 | transformers | ||
uclanlp/visualbert-vcr-coco-pre | 2021-05-31T11:27:41.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 12 | transformers | ||
uclanlp/visualbert-vcr-pre | 2021-05-31T11:29:46.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 6 | transformers | ||
uclanlp/visualbert-vcr | 2021-05-31T11:12:33.000Z | [
"pytorch",
"visual_bert",
"multiple-choice",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 47 | transformers | ||
uclanlp/visualbert-vqa-coco-pre | 2021-05-31T11:34:13.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 78 | transformers | ||
uclanlp/visualbert-vqa-pre | 2021-05-31T11:37:02.000Z | [
"pytorch",
"visual_bert",
"pretraining",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 8 | transformers | ||
uclanlp/visualbert-vqa | 2021-05-31T11:32:07.000Z | [
"pytorch",
"visual_bert",
"question-answering",
"transformers"
] | question-answering | [
".gitattributes",
"config.json",
"pytorch_model.bin"
] | uclanlp | 120 | transformers | |
ucsky/model_test | 2021-04-29T15:34:27.000Z | [] | [
".gitattributes"
] | ucsky | 0 | |||
udolda/rugpt3ForDiplom | 2021-06-12T13:41:15.000Z | [] | [
".gitattributes"
] | udolda | 0 | |||
uer/albert-base-chinese-cluecorpussmall | 2021-06-15T08:16:23.000Z | [
"pytorch",
"tf",
"albert",
"masked-lm",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 86 | transformers | |
uer/bart-base-chinese-cluecorpussmall | 2021-06-03T03:52:20.000Z | [
"pytorch",
"bart",
"seq2seq",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.txt"
] | uer | 1,162 | transformers | |
uer/chinese_roberta_L-10_H-128 | 2021-05-20T08:21:59.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 21 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-10_H-256 | 2021-05-20T08:22:21.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 10 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-10_H-512 | 2021-05-20T08:22:52.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 14 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-10_H-768 | 2021-05-20T08:23:39.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 13 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-12_H-128 | 2021-05-20T08:24:07.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 28 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-12_H-256 | 2021-05-20T08:24:30.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 7 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-12_H-512 | 2021-05-20T08:25:04.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 35 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-12_H-768 | 2021-05-20T08:26:09.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 600 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-2_H-128 | 2021-05-20T08:26:47.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 636 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-2_H-256 | 2021-05-20T08:27:07.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 251 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-2_H-512 | 2021-05-20T08:27:31.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 162 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-2_H-768 | 2021-05-20T08:27:58.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 19 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-4_H-128 | 2021-05-20T08:28:20.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 13 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-4_H-256 | 2021-05-20T08:28:42.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 160 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-4_H-512 | 2021-05-20T08:29:08.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 144 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-4_H-768 | 2021-05-20T08:29:43.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 34 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-6_H-128 | 2021-05-20T08:30:06.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 13 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-6_H-256 | 2021-05-20T08:30:30.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 18 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-6_H-512 | 2021-05-20T08:30:57.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 84 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-6_H-768 | 2021-05-20T08:31:37.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 28 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-8_H-128 | 2021-05-20T08:32:17.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 14 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-8_H-256 | 2021-05-20T08:32:38.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 10 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-8_H-512 | 2021-05-20T08:33:05.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 165 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/chinese_roberta_L-8_H-768 | 2021-05-20T08:33:46.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"arxiv:1908.08962",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 12 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "北京是[MASK]国的首都。"
---
# Chinese RoBERTa Miniatures
## Model description
This is the set of 24 Chinese RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
[Turc et al.](https://arxiv.org/abs/1908.08962) 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](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | H=128 | H=256 | H=512 | H=768 |
| -------- | :-----------------------: | :-----------------------: | :-------------------------: | :-------------------------: |
| **L=2** | [**2/128 (Tiny)**][2_128] | [2/256][2_256] | [2/512][2_512] | [2/768][2_768] |
| **L=4** | [4/128][4_128] | [**4/256 (Mini)**][4_256] | [**4/512 (Small)**][4_512] | [4/768][4_768] |
| **L=6** | [6/128][6_128] | [6/256][6_256] | [6/512][6_512] | [6/768][6_768] |
| **L=8** | [8/128][8_128] | [8/256][8_256] | [**8/512 (Medium)**][8_512] | [8/768][8_768] |
| **L=10** | [10/128][10_128] | [10/256][10_256] | [10/512][10_512] | [10/768][10_768] |
| **L=12** | [12/128][12_128] | [12/256][12_256] | [12/512][12_512] | [**12/768 (Base)**][12_768] |
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):
```python
>>> 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:
```python
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:
```python
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](https://github.com/CLUEbenchmark/CLUECorpus2020/) 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](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). 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
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 --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
```
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 --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
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/chinese_roberta_L-2_H-128
[2_256]:https://huggingface.co/uer/chinese_roberta_L-2_H-256
[2_512]:https://huggingface.co/uer/chinese_roberta_L-2_H-512
[2_768]:https://huggingface.co/uer/chinese_roberta_L-2_H-768
[4_128]:https://huggingface.co/uer/chinese_roberta_L-4_H-128
[4_256]:https://huggingface.co/uer/chinese_roberta_L-4_H-256
[4_512]:https://huggingface.co/uer/chinese_roberta_L-4_H-512
[4_768]:https://huggingface.co/uer/chinese_roberta_L-4_H-768
[6_128]:https://huggingface.co/uer/chinese_roberta_L-6_H-128
[6_256]:https://huggingface.co/uer/chinese_roberta_L-6_H-256
[6_512]:https://huggingface.co/uer/chinese_roberta_L-6_H-512
[6_768]:https://huggingface.co/uer/chinese_roberta_L-6_H-768
[8_128]:https://huggingface.co/uer/chinese_roberta_L-8_H-128
[8_256]:https://huggingface.co/uer/chinese_roberta_L-8_H-256
[8_512]:https://huggingface.co/uer/chinese_roberta_L-8_H-512
[8_768]:https://huggingface.co/uer/chinese_roberta_L-8_H-768
[10_128]:https://huggingface.co/uer/chinese_roberta_L-10_H-128
[10_256]:https://huggingface.co/uer/chinese_roberta_L-10_H-256
[10_512]:https://huggingface.co/uer/chinese_roberta_L-10_H-512
[10_768]:https://huggingface.co/uer/chinese_roberta_L-10_H-768
[12_128]:https://huggingface.co/uer/chinese_roberta_L-12_H-128
[12_256]:https://huggingface.co/uer/chinese_roberta_L-12_H-256
[12_512]:https://huggingface.co/uer/chinese_roberta_L-12_H-512
[12_768]:https://huggingface.co/uer/chinese_roberta_L-12_H-768 |
uer/gpt2-base-chinese-cluecorpussmall | 2021-05-23T13:19:22.000Z | [
"pytorch",
"tf",
"gpt2",
"lm-head",
"causal-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 643 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "米饭是一种用稻米与水煮成的食物"
---
<br>
**The GPT-2-base model with sequence length 1024 is available in [here](https://huggingface.co/uer/gpt2-chinese-cluecorpussmall). The model in this repository only supports sequence length 512. We will delete this repository in the near future.**
<br>
# Chinese GPT2 Model
## Model description
The model is used to generate Chinese texts. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-base-chinese-cluecorpussmall](https://huggingface.co/uer/gpt2-base-chinese-cluecorpussmall).
## How to use
You can use the model directly with a pipeline for text generation:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-base-chinese-cluecorpussmall")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-base-chinese-cluecorpussmall")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("这是很久之前的事情了", max_length=100, do_sample=True)
[{'generated_text': '这是很久之前的事情了 ！ 这 件 事 情 之 后 我 每 天 都 问 自 己 ， 对 未 来 的 影 响 是 什 么 ？ 在 这 个 过 程 中 我 一 直 提 高 自 己 的 理 论 和 实 践 能 力 ， 比 如 说 ， 我 们 现 在 有 很 多 很 多 的 投 资 行 为 可 以 赚 钱 ， 在 美 国 有 很 多 交 易 行 为 ， 是 一 个 比 较 灵 活 的 模'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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.
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--seq_length 128 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--vocab_path models/google_zh_vocab.txt \
--config_path models/bert_base_config.json \
--output_model_path models/cluecorpussmall_gpt2_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq512_dataset.pt \
--seq_length 512 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_gpt2_seq128_model.bin-1000000 \
--vocab_path models/google_zh_vocab.txt \
--config_path models/bert_base_config.json \
--output_model_path models/cluecorpussmall_gpt2_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path cluecorpussmall_gpt2_seq512_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
``` |
uer/gpt2-chinese-ancient | 2021-05-23T13:20:09.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"Chinese",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 258 | transformers | ---
language: Chinese
widget:
- text: "[CLS]当是时"
---
# Chinese Ancient GPT2 Model
## Model description
The model is used to generate ancient Chinese. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-ancient](https://huggingface.co/uer/gpt2-chinese-ancient)
## How to use
You can use the model directly with a pipeline for text generation:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-ancient")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-ancient")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("当是时", max_length=100, do_sample=True)
[{'generated_text': '[CLS]当是时 所 议 者 不 为 无 据 ， 况 亦 在 之 列 乎 ？ 然 则 今 日 之 事 ， 所 当 思 者 在 何 ？ 欲 求 国 是 于 天 下 ， 莫 在 于 得 人 。 臣 以 为 求 人 之 法 ， 不 在 多 用 官 一 途 。 诚 使 得 才 者 众 ， 人 才 者 优 ， 则 治 所 当 得 ， 而 不 事 于 官 者 ， 人 才 乃 其 常 也 。 所 当 讲 者'}]
```
## Training data
Training data contains 3,000,000 ancient Chinese which are collected by [daizhigev20](https://github.com/garychowcmu/daizhigev20). Since part of ancient corpus has no punctuation, we used the [ancient Chinese punctuation system](https://seg.shenshen.wiki) developed by [BNU ICIP lab](http://icip.bnu.edu.cn/).
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We pre-train 500,000 steps with a sequence length of 320. We use extended vocabulary to handle out-of-vocabulary words. The Chinese character that occurs greater than or equal to 100 in ancient Chinese corpus is added to the vocabulary.
```
python3 preprocess.py --corpus_path corpora/ancient_chinese.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path ancient_chinese_dataset.pt --processes_num 16 \
--seq_length 320 --target lm
```
```
python3 pretrain.py --dataset_path ancient_chinese_dataset.pt \
--vocab_path models/google_zh_vocab.txt \
--config_path models/bert_base_config.json \
--output_model_path models/ancient_chinese_base_model.bin \
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
--total_steps 500000 --save_checkpoint_steps 100000 --report_steps 10000 \
--learning_rate 5e-4 --batch_size 32 \
--embedding word_pos --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path ancient_chinese_base_model.bin-500000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
胡韧奋,李绅,诸雨辰.基于深层语言模型的古汉语知识表示及自动断句研究[C].第十八届中国计算语言学大会（CCL 2019）.
``` |
uer/gpt2-chinese-cluecorpussmall | 2021-05-23T13:21:10.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 2,145 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "这是很久之前的事情了"
---
# Chinese GPT2 Model
## Model description
The model is used to generate Chinese texts. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-cluecorpussmall](https://huggingface.co/uer/gpt2-chinese-cluecorpussmall).
## How to use
You can use the model directly with a pipeline for text generation:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("这是很久之前的事情了", max_length=100, do_sample=True)
[{'generated_text': '这是很久之前的事情了 ， 我 曾 经 把 这 个 当 做 一 种 思 想 的 传 承 ， 或 者 是 人 生 的 回 顾 ， 当 时 我 们 是 一 个 刚 刚 加 入 的 时 候 就 想 要 加 入 他 们 ， 于 是 我 们 每 天 看 到 他 们 ， 加 上 他 们 的 各 种 不 可 思 议 的 行 为 ， 直 到 现 在 ， 我 们 的 人 生 才 完 整 起 来 。'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 1024.
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--seq_length 128 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/config.json \
--output_model_path models/cluecorpussmall_gpt2_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq1024_dataset.pt \
--seq_length 1024 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq1024_dataset.pt \
--pretrained_model_path models/cluecorpussmall_gpt2_seq128_model.bin-1000000 \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/config.json \
--output_model_path models/cluecorpussmall_gpt2_seq1024_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path cluecorpussmall_gpt2_seq1024_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
``` |
uer/gpt2-chinese-couplet | 2021-05-23T13:22:15.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"zh",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 225 | transformers | ---
language: zh
widget:
- text: "[CLS]国 色 天 香 ， 姹 紫 嫣 红 ， 碧 水 青 云 欣 共 赏 -"
---
# Chinese Couplet GPT2 Model
## Model description
The model is used to generate Chinese couplets. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-couplet][couplet].
Since the parameter skip_special_tokens is used in the pipelines.py, special tokens such as [SEP], [UNK] will be deleted, the output results of Hosted inference API (right) may not be properly displayed..
## How to use
You can use the model directly with a pipeline for text generation:
When the parameter skip_special_tokens is True:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-couplet")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-couplet")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("[CLS]丹 枫 江 冷 人 初 去 -", max_length=25, do_sample=True)
[{'generated_text': '[CLS]丹 枫 江 冷 人 初 去 - 黄 叶 声 从 天 外 来 阅 旗'}]
```
When the parameter skip_special_tokens is False:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-couplet")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-couplet")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("[CLS]丹 枫 江 冷 人 初 去 -", max_length=25, do_sample=True)
[{'generated_text': '[CLS]丹 枫 江 冷 人 初 去 - 黄 叶 声 我 酒 不 辞 [SEP] [SEP] [SEP] [SEP] [SEP] [SEP] [SEP] [SEP] [SEP]'}]
```
## Training data
Training data contains 700,000 Chinese couplets which are collected by [couplet-clean-dataset](https://github.com/v-zich/couplet-clean-dataset).
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We pre-train 25,000 steps with a sequence length of 64.
```
python3 preprocess.py --corpus_path corpora/couplet.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path couplet_dataset.pt --processes_num 16 \
--seq_length 64 --target lm
```
```
python3 pretrain.py --dataset_path couplet_dataset.pt \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/config.json \
--output_model_path models/couplet_gpt2_model.bin \
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
--total_steps 25000 --save_checkpoint_steps 5000 --report_steps 1000 \
--learning_rate 5e-4 --batch_size 64 \
--embedding word_pos --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path couplet_gpt2_model.bin-25000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
```
[couplet]: https://huggingface.co/uer/gpt2-chinese-couplet |
uer/gpt2-chinese-lyric | 2021-05-23T13:23:13.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"Chinese",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 237 | transformers | ---
language: Chinese
widget:
- text: "最美的不是下雨天，是曾与你躲过雨的屋檐"
---
# Chinese GPT2 Lyric Model
## Model description
The model is used to generate Chinese lyrics. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-lyric](https://huggingface.co/uer/gpt2-chinese-lyric)
## How to use
You can use the model directly with a pipeline for text generation:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-lyric")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-lyric")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("最美的不是下雨天，是曾与你躲过雨的屋檐", max_length=100, do_sample=True)
[{'generated_text': '最美的不是下雨天，是曾与你躲过雨的屋檐 ， 下 课 铃 声 响 起 的 瞬 间 ， 我 们 的 笑 脸 ， 有 太 多 回 忆 在 浮 现 ， 是 你 总 在 我 身 边 ， 不 知 道 会 不 会 再 见 ， 从 现 在 开 始 到 永 远 ， 想 说 的 语 言 凝 结 成 一 句 ， 不 管 我 们 是 否 能 够 兑 现 ， 想 说 的 语 言 凝 结'}]
```
## Training data
Training data contains 150,000 Chinese lyrics which are collected by [Chinese-Lyric-Corpus](https://github.com/gaussic/Chinese-Lyric-Corpus) and [MusicLyricChatbot](https://github.com/liuhuanyong/MusicLyricChatbot).
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We pre-train 100,000 steps with a sequence length of 512 on the basis of the pre-trained model [gpt2-base-chinese-cluecorpussmall](https://huggingface.co/uer/gpt2-base-chinese-cluecorpussmall)
```
python3 preprocess.py --corpus_path corpora/lyric.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path lyric_dataset.pt --processes_num 32 \
--seq_length 512 --target lm
```
```
python3 pretrain.py --dataset_path lyric_dataset.pt \
--pretrained_model_path models/cluecorpussmall_gpt2_seq1024_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/config.json \
--output_model_path models/lyric_gpt2_model.bin \
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
--total_steps 100000 --save_checkpoint_steps 10000 --report_steps 5000 \
--learning_rate 5e-5 --batch_size 64 \
--embedding word_pos --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path lyric_gpt2_model.bin-100000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
``` |
uer/gpt2-chinese-poem | 2021-05-23T13:24:24.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"zh",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 1,686 | transformers | ---
language: zh
widget:
- text: "[CLS] 万 叠 春 山 积 雨 晴 ，"
- text: "[CLS] 大 漠"
---
# Chinese Poem GPT2 Model
## Model description
The model is used to generate Chinese ancient poems. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-poem][poem].
Since the parameter skip_special_tokens is used in the pipelines.py, special tokens such as [SEP], [UNK] will be deleted, the output results of Hosted inference API (right) may not be properly displayed.
## How to use
You can use the model directly with a pipeline for text generation:
When the parameter skip_special_tokens is True:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel,TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-poem")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("[CLS]梅 山 如 积 翠 ，", max_length=50, do_sample=True)
[{'generated_text': '[CLS]梅 山 如 积 翠 ， 丛 竹 隠 疏 花 。 水 影 落 寒 濑 ， 竹 声 随 暮 鸦 。 茅 茨 数 间 屋 ， 烟 火 两 三 家 。 安 得 携 琴 酒 ， 相 逢 烟 雨 赊 。 向 湖 边 过 ， 偏 怜 雪 里 看 。 浮 峦 如 画 出 ， 远 树 与 天 连 。 月 上 僧 房 静 ， 风 回 萤 火 寒 。 幽 情 何 可 写 ， 赖 有 子 期 弹 。 棠 真'}]
```
When the parameter skip_special_tokens is False:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel,TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-poem")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("[CLS]梅 山 如 积 翠 ，", max_length=100, do_sample=True)
[{'generated_text': '[CLS]梅 山 如 积 翠 ， 秀 出 何 其 雄 。 矫 矫 云 间 质 ， 映 日 生 玲 珑 。 根 大 乱 石 结 ， 枝 高 青 云 蒙 。 常 因 风 露 晚 ， 隠 映 瑶 台 中 。 忽 闻 山 石 裂 ， 万 里 吹 天 风 。 又 觉 此 身 高 ， 迥 出 凡 境 空 。 清 影 落 潭 水 ， 暗 香 来 逈 峰 。 却 寻 白 太 白 ， 月 影 摇 江 东 。 [SEP] 而 非'}]
```
## Training data
Training data contains 800,000 Chinese ancient poems which are collected by [chinese-poetry](https://github.com/chinese-poetry/chinese-poetry) and [Poetry](https://github.com/Werneror/Poetry) projects.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We pre-train 200,000 steps with a sequence length of 128. We use extended vocabulary to handle out-of-vocabulary words. The Chinese character that occurs greater than or equal to 100 in poem corpus is added to the vocabulary.
```
python3 preprocess.py --corpus_path corpora/poem.txt \
--vocab_path models/poem_zh_vocab.txt \
--dataset_path poem_dataset.pt --processes_num 16 \
--seq_length 128 --target lm
```
```
python3 pretrain.py --dataset_path poem_dataset.pt \
--vocab_path models/poem_zh_vocab.txt \
--config_path models/gpt2/config.json \
--output_model_path models/poem_gpt2_model.bin \
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
--total_steps 200000 --save_checkpoint_steps 50000 --report_steps 1000 \
--learning_rate 5e-4 --batch_size 64 \
--embedding word_pos --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path poem_gpt2_base_model.bin-200000 \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
```
[poem]: https://huggingface.co/uer/gpt2-chinese-poem |
uer/gpt2-distil-chinese-cluecorpussmall | 2021-05-23T13:25:16.000Z | [
"pytorch",
"tf",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"Chinese",
"dataset:CLUECorpusSmall",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 193 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "米饭是一种用稻米与水煮成的食物"
---
# Chinese GPT2-distil Model
## Model description
The model is used to generate Chinese texts. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-distil-chinese-cluecorpussmall](https://huggingface.co/uer/gpt2-distil-chinese-cluecorpussmall). The model is called GPT2-distil because the configuration of model follows [distilgpt2](https://huggingface.co/distilgpt2), which has 6 layers, 768 dimension, and 12 heads. The pre-training does not involve the supervision of larger models.
## How to use
You can use the model directly with a pipeline for text generation:
```python
>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-distil-chinese-cluecorpussmall")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-distil-chinese-cluecorpussmall")
>>> text_generator = TextGenerationPipeline(model, tokenizer)
>>> text_generator("这是很久之前的事情了", max_length=100, do_sample=True)
[{'generated_text': '这是很久之前的事情了 。 我 现 在 想 起 来 就 让 自 己 很 伤 心 ， 很 失 望 。 我 现 在 想 到 ， 我 觉 得 大 多 数 人 的 生 活 比 我 的 生 命 还 要 重 要 ， 对 一 些 事 情 的 看 法 ， 对 一 些 人 的 看 法 ， 都 是 在 发 泄 。 但 是 ， 我 们 的 生 活 是 需 要 一 个 信 用 体 系 的 。 我 不 知'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 1024.
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--seq_length 128 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq128_dataset.pt \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/distil_config.json \
--output_model_path models/cluecorpussmall_gpt2_distil_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_vocab.txt \
--dataset_path cluecorpussmall_lm_seq1024_dataset.pt \
--seq_length 1024 --processes_num 32 --target lm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_lm_seq1024_dataset.pt \
--pretrained_model_path models/cluecorpussmall_gpt2_distil_seq128_model.bin-1000000 \
--vocab_path models/google_zh_vocab.txt \
--config_path models/gpt2/distil_config.json \
--output_model_path models/cluecorpussmall_gpt2_distil_seq1024_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 --remove_embedding_layernorm \
--encoder transformer --mask causal --layernorm_positioning pre \
--target lm --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path cluecorpussmall_gpt2_distil_seq1024_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 6
```
### BibTeX entry and citation info
```
@article{radford2019language,
title={Language Models are Unsupervised Multitask Learners},
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
year={2019}
}
@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}
}
``` |
uer/pegasus-base-chinese-cluecorpussmall | 2021-06-02T07:12:09.000Z | [
"pytorch",
"pegasus",
"seq2seq",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.txt"
] | uer | 58 | transformers | |
uer/roberta-base-chinese-extractive-qa | 2021-05-20T08:34:38.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"question-answering",
"zh",
"transformers"
] | question-answering | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 2,577 | transformers | ---
language: zh
widget:
- text: "著名诗歌《假如生活欺骗了你》的作者是"
context: "普希金从那里学习人民的语言，吸取了许多有益的养料，这一切对普希金后来的创作产生了很大的影响。这两年里，普希金创作了不少优秀的作品，如《囚徒》、《致大海》、《致凯恩》和《假如生活欺骗了你》等几十首抒情诗，叙事诗《努林伯爵》，历史剧《鲍里斯·戈都诺夫》，以及《叶甫盖尼·奥涅金》前六章。"
---
# Chinese RoBERTa-Base Model for QA
## Model description
The model is used for extractive question answering. You can download the model from the link [roberta-base-chinese-extractive-qa](https://huggingface.co/uer/roberta-base-chinese-extractive-qa).
## How to use
You can use the model directly with a pipeline for extractive question answering:
```python
>>> from transformers import AutoModelForQuestionAnswering,AutoTokenizer,pipeline
>>> model = AutoModelForQuestionAnswering.from_pretrained('uer/roberta-base-chinese-extractive-qa')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-chinese-extractive-qa')
>>> QA = pipeline('question-answering', model=model, tokenizer=tokenizer)
>>> QA_input = {'question': "著名诗歌《假如生活欺骗了你》的作者是",'context': "普希金从那里学习人民的语言，吸取了许多有益的养料，这一切对普希金后来的创作产生了很大的影响。这两年里，普希金创作了不少优秀的作品，如《囚徒》、《致大海》、《致凯恩》和《假如生活欺骗了你》等几十首抒情诗，叙事诗《努林伯爵》，历史剧《鲍里斯·戈都诺夫》，以及《叶甫盖尼·奥涅金》前六章。"}
>>> QA(QA_input)
{'score': 0.9766426682472229, 'start': 0, 'end': 3, 'answer': '普希金'}
```
## Training data
Training data comes from three sources: [cmrc2018](https://github.com/ymcui/cmrc2018), [webqa](https://spaces.ac.cn/archives/4338), and [laisi](https://www.kesci.com/home/competition/5d142d8cbb14e6002c04e14a/content/0). We only use the train set of three datasets.
## Training procedure
The model is fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud TI-ONE](https://cloud.tencent.com/product/tione/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved.
```
python3 run_cmrc.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path extractive_qa.json \
--dev_path datasets/cmrc2018/dev.json \
--output_model_path models/extractive_qa_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the fine-tuned model into Huggingface's format:
```
python3 scripts/convert_bert_extractive_qa_from_uer_to_huggingface.py --input_model_path extractive_qa_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### 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}
}
``` |
uer/roberta-base-finetuned-chinanews-chinese | 2021-05-20T08:36:07.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"zh",
"arxiv:1909.05658",
"arxiv:1708.02657",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 652 | transformers | ---
language: zh
widget:
- text: "北京上个月召开了两会"
---
# Chinese RoBERTa-Base Models for Text Classification
## Model description
This is the set of 5 Chinese RoBERTa-Base classification models fine-tuned by [UER-py](https://arxiv.org/abs/1909.05658). You can download the 5 Chinese RoBERTa-Base classification models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the links below:
| Dataset | Link |
| :-----------: | :-------------------------------------------------------: |
| **JD full** | [**roberta-base-finetuned-jd-full-chinese**][jd_full] |
| **JD binary** | [**roberta-base-finetuned-jd-binary-chinese**][jd_binary] |
| **Dianping** | [**roberta-base-finetuned-dianping-chinese**][dianping] |
| **Ifeng** | [**roberta-base-finetuned-ifeng-chinese**][ifeng] |
| **Chinanews** | [**roberta-base-finetuned-chinanews-chinese**][chinanews] |
## How to use
You can use this model directly with a pipeline for text classification (take the case of roberta-base-finetuned-chinanews-chinese):
```python
>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
[{'label': 'mainland China politics', 'score': 0.7211663722991943}]
```
## Training data
5 Chinese text classification datasets are used. JD full, JD binary, and Dianping datasets consist of user reviews of different sentiment polarities. Ifeng and Chinanews consist of first paragraphs of news articles of different topic classes. They are collected by [Glyph](https://github.com/zhangxiangxiao/glyph) project and more details are discussed in corresponding [paper](https://arxiv.org/abs/1708.02657).
## Training procedure
Models are fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved. We use the same hyper-parameters on different models.
Taking the case of roberta-base-finetuned-chinanews-chinese
```
python3 run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/glyph/chinanews/train.tsv \
--dev_path datasets/glyph/chinanews/dev.tsv \
--output_model_path models/chinanews_classifier_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{zhang2017encoding,
title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
author={Zhang, Xiang and LeCun, Yann},
journal={arXiv preprint arXiv:1708.02657},
year={2017}
}
@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}
}
```
[jd_full]:https://huggingface.co/uer/roberta-base-finetuned-jd-full-chinese
[jd_binary]:https://huggingface.co/uer/roberta-base-finetuned-jd-binary-chinese
[dianping]:https://huggingface.co/uer/roberta-base-finetuned-dianping-chinese
[ifeng]:https://huggingface.co/uer/roberta-base-finetuned-ifeng-chinese
[chinanews]:https://huggingface.co/uer/roberta-base-finetuned-chinanews-chinese |
uer/roberta-base-finetuned-cluener2020-chinese | 2021-05-20T08:37:15.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"token-classification",
"zh",
"transformers"
] | token-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 536 | transformers | ---
language: zh
widget:
- text: "江苏警方通报特斯拉冲进店铺"
---
# Chinese RoBERTa-Base Model for NER
## Model description
The model is used for named entity recognition. You can download the model either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the link [roberta-base-finetuned-cluener2020-chinese](https://huggingface.co/uer/roberta-base-finetuned-cluener2020-chinese).
## How to use
You can use this model directly with a pipeline for token classification :
```python
>>> from transformers import AutoModelForTokenClassification,AutoTokenizer,pipeline
>>> model = AutoModelForTokenClassification.from_pretrained('uer/roberta-base-finetuned-cluener2020-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-cluener2020-chinese')
>>> ner = pipeline('ner', model=model, tokenizer=tokenizer)
>>> ner("江苏警方通报特斯拉冲进店铺")
[
{'word': '江', 'score': 0.49153077602386475, 'entity': 'B-address', 'index': 1, 'start': 0, 'end': 1},
{'word': '苏', 'score': 0.6319217681884766, 'entity': 'I-address', 'index': 2, 'start': 1, 'end': 2},
{'word': '特', 'score': 0.5912262797355652, 'entity': 'B-company', 'index': 7, 'start': 6, 'end': 7},
{'word': '斯', 'score': 0.69145667552948, 'entity': 'I-company', 'index': 8, 'start': 7, 'end': 8},
{'word': '拉', 'score': 0.7054660320281982, 'entity': 'I-company', 'index': 9, 'start': 8, 'end': 9}
]
```
## Training data
[CLUENER2020](https://github.com/CLUEbenchmark/CLUENER2020) is used as training data. We only use the train set of the dataset.
## Training procedure
The model is fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune five epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved.
```
python3 run_ner.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/cluener2020/train.tsv \
--dev_path datasets/cluener2020/dev.tsv \
--label2id_path datasets/cluener2020/label2id.json \
--output_model_path models/cluener2020_ner_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 5 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_token_classification_from_uer_to_huggingface.py --input_model_path models/cluener2020_ner_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{xu2020cluener2020,
title={CLUENER2020: Fine-grained Name Entity Recognition for Chinese},
author={Xu, Liang and Dong, Qianqian and Yu, Cong and Tian, Yin and Liu, Weitang and Li, Lu and Zhang, Xuanwei},
journal={arXiv preprint arXiv:2001.04351},
year={2020}
}
@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}
}
``` |
uer/roberta-base-finetuned-dianping-chinese | 2021-05-20T08:38:11.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"zh",
"arxiv:1909.05658",
"arxiv:1708.02657",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 696 | transformers | ---
language: zh
widget:
- text: "这本书真的很不错"
---
# Chinese RoBERTa-Base Models for Text Classification
## Model description
This is the set of 5 Chinese RoBERTa-Base classification models fine-tuned by [UER-py](https://arxiv.org/abs/1909.05658). You can download the 5 Chinese RoBERTa-Base classification models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the links below:
| Dataset | Link |
| :-----------: | :-------------------------------------------------------: |
| **JD full** | [**roberta-base-finetuned-jd-full-chinese**][jd_full] |
| **JD binary** | [**roberta-base-finetuned-jd-binary-chinese**][jd_binary] |
| **Dianping** | [**roberta-base-finetuned-dianping-chinese**][dianping] |
| **Ifeng** | [**roberta-base-finetuned-ifeng-chinese**][ifeng] |
| **Chinanews** | [**roberta-base-finetuned-chinanews-chinese**][chinanews] |
## How to use
You can use this model directly with a pipeline for text classification (take the case of roberta-base-finetuned-chinanews-chinese):
```python
>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
[{'label': 'mainland China politics', 'score': 0.7211663722991943}]
```
## Training data
5 Chinese text classification datasets are used. JD full, JD binary, and Dianping datasets consist of user reviews of different sentiment polarities. Ifeng and Chinanews consist of first paragraphs of news articles of different topic classes. They are collected by [Glyph](https://github.com/zhangxiangxiao/glyph) project and more details are discussed in corresponding [paper](https://arxiv.org/abs/1708.02657).
## Training procedure
Models are fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved. We use the same hyper-parameters on different models.
Taking the case of roberta-base-finetuned-chinanews-chinese
```
python3 run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/glyph/chinanews/train.tsv \
--dev_path datasets/glyph/chinanews/dev.tsv \
--output_model_path models/chinanews_classifier_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{zhang2017encoding,
title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
author={Zhang, Xiang and LeCun, Yann},
journal={arXiv preprint arXiv:1708.02657},
year={2017}
}
@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}
}
```
[jd_full]:https://huggingface.co/uer/roberta-base-finetuned-jd-full-chinese
[jd_binary]:https://huggingface.co/uer/roberta-base-finetuned-jd-binary-chinese
[dianping]:https://huggingface.co/uer/roberta-base-finetuned-dianping-chinese
[ifeng]:https://huggingface.co/uer/roberta-base-finetuned-ifeng-chinese
[chinanews]:https://huggingface.co/uer/roberta-base-finetuned-chinanews-chinese |
uer/roberta-base-finetuned-ifeng-chinese | 2021-05-20T08:39:05.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"zh",
"arxiv:1909.05658",
"arxiv:1708.02657",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 59 | transformers | ---
language: zh
widget:
- text: "北京上个月召开了两会"
---
# Chinese RoBERTa-Base Models for Text Classification
## Model description
This is the set of 5 Chinese RoBERTa-Base classification models fine-tuned by [UER-py](https://arxiv.org/abs/1909.05658). You can download the 5 Chinese RoBERTa-Base classification models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the links below:
| Dataset | Link |
| :-----------: | :-------------------------------------------------------: |
| **JD full** | [**roberta-base-finetuned-jd-full-chinese**][jd_full] |
| **JD binary** | [**roberta-base-finetuned-jd-binary-chinese**][jd_binary] |
| **Dianping** | [**roberta-base-finetuned-dianping-chinese**][dianping] |
| **Ifeng** | [**roberta-base-finetuned-ifeng-chinese**][ifeng] |
| **Chinanews** | [**roberta-base-finetuned-chinanews-chinese**][chinanews] |
## How to use
You can use this model directly with a pipeline for text classification (take the case of roberta-base-finetuned-chinanews-chinese):
```python
>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
[{'label': 'mainland China politics', 'score': 0.7211663722991943}]
```
## Training data
5 Chinese text classification datasets are used. JD full, JD binary, and Dianping datasets consist of user reviews of different sentiment polarities. Ifeng and Chinanews consist of first paragraphs of news articles of different topic classes. They are collected by [Glyph](https://github.com/zhangxiangxiao/glyph) project and more details are discussed in corresponding [paper](https://arxiv.org/abs/1708.02657).
## Training procedure
Models are fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved. We use the same hyper-parameters on different models.
Taking the case of roberta-base-finetuned-chinanews-chinese
```
python3 run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/glyph/chinanews/train.tsv \
--dev_path datasets/glyph/chinanews/dev.tsv \
--output_model_path models/chinanews_classifier_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{zhang2017encoding,
title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
author={Zhang, Xiang and LeCun, Yann},
journal={arXiv preprint arXiv:1708.02657},
year={2017}
}
@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}
}
```
[jd_full]:https://huggingface.co/uer/roberta-base-finetuned-jd-full-chinese
[jd_binary]:https://huggingface.co/uer/roberta-base-finetuned-jd-binary-chinese
[dianping]:https://huggingface.co/uer/roberta-base-finetuned-dianping-chinese
[ifeng]:https://huggingface.co/uer/roberta-base-finetuned-ifeng-chinese
[chinanews]:https://huggingface.co/uer/roberta-base-finetuned-chinanews-chinese |
uer/roberta-base-finetuned-jd-binary-chinese | 2021-05-20T08:40:01.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"zh",
"arxiv:1909.05658",
"arxiv:1708.02657",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 149 | transformers | ---
language: zh
widget:
- text: "这本书真的很不错"
---
# Chinese RoBERTa-Base Models for Text Classification
## Model description
This is the set of 5 Chinese RoBERTa-Base classification models fine-tuned by [UER-py](https://arxiv.org/abs/1909.05658). You can download the 5 Chinese RoBERTa-Base classification models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the links below:
| Dataset | Link |
| :-----------: | :-------------------------------------------------------: |
| **JD full** | [**roberta-base-finetuned-jd-full-chinese**][jd_full] |
| **JD binary** | [**roberta-base-finetuned-jd-binary-chinese**][jd_binary] |
| **Dianping** | [**roberta-base-finetuned-dianping-chinese**][dianping] |
| **Ifeng** | [**roberta-base-finetuned-ifeng-chinese**][ifeng] |
| **Chinanews** | [**roberta-base-finetuned-chinanews-chinese**][chinanews] |
## How to use
You can use this model directly with a pipeline for text classification (take the case of roberta-base-finetuned-chinanews-chinese):
```python
>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
[{'label': 'mainland China politics', 'score': 0.7211663722991943}]
```
## Training data
5 Chinese text classification datasets are used. JD full, JD binary, and Dianping datasets consist of user reviews of different sentiment polarities. Ifeng and Chinanews consist of first paragraphs of news articles of different topic classes. They are collected by [Glyph](https://github.com/zhangxiangxiao/glyph) project and more details are discussed in corresponding [paper](https://arxiv.org/abs/1708.02657).
## Training procedure
Models are fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved. We use the same hyper-parameters on different models.
Taking the case of roberta-base-finetuned-chinanews-chinese
```
python3 run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/glyph/chinanews/train.tsv \
--dev_path datasets/glyph/chinanews/dev.tsv \
--output_model_path models/chinanews_classifier_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{zhang2017encoding,
title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
author={Zhang, Xiang and LeCun, Yann},
journal={arXiv preprint arXiv:1708.02657},
year={2017}
}
@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}
}
```
[jd_full]:https://huggingface.co/uer/roberta-base-finetuned-jd-full-chinese
[jd_binary]:https://huggingface.co/uer/roberta-base-finetuned-jd-binary-chinese
[dianping]:https://huggingface.co/uer/roberta-base-finetuned-dianping-chinese
[ifeng]:https://huggingface.co/uer/roberta-base-finetuned-ifeng-chinese
[chinanews]:https://huggingface.co/uer/roberta-base-finetuned-chinanews-chinese |
uer/roberta-base-finetuned-jd-full-chinese | 2021-05-20T08:40:53.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"zh",
"arxiv:1909.05658",
"arxiv:1708.02657",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 138 | transformers | ---
language: zh
widget:
- text: "这本书真的很不错"
---
# Chinese RoBERTa-Base Models for Text Classification
## Model description
This is the set of 5 Chinese RoBERTa-Base classification models fine-tuned by [UER-py](https://arxiv.org/abs/1909.05658). You can download the 5 Chinese RoBERTa-Base classification models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo) (in UER-py format), or via HuggingFace from the links below:
| Dataset | Link |
| :-----------: | :-------------------------------------------------------: |
| **JD full** | [**roberta-base-finetuned-jd-full-chinese**][jd_full] |
| **JD binary** | [**roberta-base-finetuned-jd-binary-chinese**][jd_binary] |
| **Dianping** | [**roberta-base-finetuned-dianping-chinese**][dianping] |
| **Ifeng** | [**roberta-base-finetuned-ifeng-chinese**][ifeng] |
| **Chinanews** | [**roberta-base-finetuned-chinanews-chinese**][chinanews] |
## How to use
You can use this model directly with a pipeline for text classification (take the case of roberta-base-finetuned-chinanews-chinese):
```python
>>> from transformers import AutoModelForSequenceClassification,AutoTokenizer,pipeline
>>> model = AutoModelForSequenceClassification.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> tokenizer = AutoTokenizer.from_pretrained('uer/roberta-base-finetuned-chinanews-chinese')
>>> text_classification = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
>>> text_classification("北京上个月召开了两会")
[{'label': 'mainland China politics', 'score': 0.7211663722991943}]
```
## Training data
5 Chinese text classification datasets are used. JD full, JD binary, and Dianping datasets consist of user reviews of different sentiment polarities. Ifeng and Chinanews consist of first paragraphs of news articles of different topic classes. They are collected by [Glyph](https://github.com/zhangxiangxiao/glyph) project and more details are discussed in corresponding [paper](https://arxiv.org/abs/1708.02657).
## Training procedure
Models are fine-tuned by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We fine-tune three epochs with a sequence length of 512 on the basis of the pre-trained model [chinese_roberta_L-12_H-768](https://huggingface.co/uer/chinese_roberta_L-12_H-768). At the end of each epoch, the model is saved when the best performance on development set is achieved. We use the same hyper-parameters on different models.
Taking the case of roberta-base-finetuned-chinanews-chinese
```
python3 run_classifier.py --pretrained_model_path models/cluecorpussmall_roberta_base_seq512_model.bin-250000 \
--vocab_path models/google_zh_vocab.txt \
--train_path datasets/glyph/chinanews/train.tsv \
--dev_path datasets/glyph/chinanews/dev.tsv \
--output_model_path models/chinanews_classifier_model.bin \
--learning_rate 3e-5 --batch_size 32 --epochs_num 3 --seq_length 512 \
--embedding word_pos_seg --encoder transformer --mask fully_visible
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_bert_text_classification_from_uer_to_huggingface.py --input_model_path models/chinanews_classifier_model.bin \
--output_model_path pytorch_model.bin \
--layers_num 12
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{liu2019roberta,
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},
year={2019}
}
@article{zhang2017encoding,
title={Which encoding is the best for text classification in chinese, english, japanese and korean?},
author={Zhang, Xiang and LeCun, Yann},
journal={arXiv preprint arXiv:1708.02657},
year={2017}
}
@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}
}
```
[jd_full]:https://huggingface.co/uer/roberta-base-finetuned-jd-full-chinese
[jd_binary]:https://huggingface.co/uer/roberta-base-finetuned-jd-binary-chinese
[dianping]:https://huggingface.co/uer/roberta-base-finetuned-dianping-chinese
[ifeng]:https://huggingface.co/uer/roberta-base-finetuned-ifeng-chinese
[chinanews]:https://huggingface.co/uer/roberta-base-finetuned-chinanews-chinese |
uer/roberta-base-word-chinese-cluecorpussmall | 2021-05-20T08:42:28.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"zh",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tf_model.h5",
"tokenizer_config.json"
] | uer | 350 | transformers | ---
language: zh
datasets: CLUECorpusSmall
widget:
- text: "最近一趟去北京的[MASK]几点发车"
---
# Chinese word-based RoBERTa Miniatures
## Model description
This is the set of 5 Chinese word-based RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
Most Chinese pre-trained weights are based on Chinese character. Compared with character-based models, word-based models are faster (because of shorter sequence length) and have better performance according to our experimental results. To this end, we released the 5 Chinese word-based RoBERTa models of different sizes. In order to facilitate users to reproduce the results, we used the publicly available corpus and word segmentation tool, and provided all training details.
Notice that the output results of Hosted inference API (right) are not properly displayed. When the predicted word has multiple characters, the single word instead of entire sentence is displayed. One can click **JSON Output** for normal output results.
You can download the 5 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | Link |
| -------- | :-----------------------: |
| **word-based RoBERTa-Tiny** | [**L=2/H=128 (Tiny)**][2_128] |
| **word-based RoBERTa-Mini** | [**L=4/H=256 (Mini)**][4_256] |
| **word-based RoBERTa-Small** | [**L=4/H=512 (Small)**][4_512] |
| **word-based RoBERTa-Medium** | [**L=8/H=512 (Medium)**][8_512] |
| **word-based RoBERTa-Base** | [**L=12/H=768 (Base)**][12_768] |
Compared with [char-based models](https://huggingface.co/uer/chinese_roberta_L-2_H-128), word-based models achieve better results in most cases. Here are scores on the devlopment set of six Chinese tasks:
| Model | Score | douban | chnsenticorp | lcqmc | tnews(CLUE) | iflytek(CLUE) | ocnli(CLUE) |
| -------------- | :---: | :----: | :----------: | :---: | :---------: | :-----------: | :---------: |
| RoBERTa-Tiny(char) | 72.3 | 83.0 | 91.4 | 81.8 | 62.0 | 55.0 | 60.3 |
| **RoBERTa-Tiny(word)** | **74.3(+2.0)** | **86.4** | **93.2** | **82.0** | **66.4** | **58.2** | **59.6** |
| RoBERTa-Mini(char) | 75.7 | 84.8 | 93.7 | 86.1 | 63.9 | 58.3 | 67.4 |
| **RoBERTa-Mini(word)** | **76.7(+1.0)** | **87.6** | **94.1** | **85.4** | **66.9** | **59.2** | **67.3** |
| RoBERTa-Small(char) | 76.8 | 86.5 | 93.4 | 86.5 | 65.1 | 59.4 | 69.7 |
| **RoBERTa-Small(word)** | **78.1(+1.3)** | **88.5** | **94.7** | **87.4** | **67.6** | **60.9** | **69.8** |
| RoBERTa-Medium(char) | 77.8 | 87.6 | 94.8 | 88.1 | 65.6 | 59.5 | 71.2 |
| **RoBERTa-Medium(word)** | **78.9(+1.1)** | **89.2** | **95.1** | **88.0** | **67.8** | **60.6** | **73.0** |
| RoBERTa-Base(char) | 79.5 | 89.1 | 95.2 | 89.2 | 67.0 | 60.9 | 75.5 |
| **RoBERTa-Base(word)** | **80.2(+0.7)** | **90.3** | **95.7** | **89.4** | **68.0** | **61.5** | **76.8** |
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 word-based RoBERTa-Medium):
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-medium-word-chinese-cluecorpussmall')
>>> unmasker("[MASK]的首都是北京。")
[
{'sequence': '中国 的首都是北京。',
'score': 0.21525809168815613,
'token': 2873,
'token_str': '中国'},
{'sequence': '北京 的首都是北京。',
'score': 0.15194718539714813,
'token': 9502,
'token_str': '北京'},
{'sequence': '我们 的首都是北京。',
'score': 0.08854265511035919,
'token': 4215,
'token_str': '我们'},
{'sequence': '美国 的首都是北京。',
'score': 0.06808705627918243,
'token': 7810,
'token_str': '美国'},
{'sequence': '日本 的首都是北京。',
'score': 0.06071401759982109,
'token': 7788,
'token_str': '日本'}
]
```
Here is how to use this model to get the features of a given text in PyTorch:
```python
from transformers import AlbertTokenizer, BertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-medium-word-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:
```python
from transformers import AlbertTokenizer, TFBertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-medium-word-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)
```
Since BertTokenizer does not support sentencepiece, AlbertTokenizer is used here.
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data. Google's [sentencepiece](https://github.com/google/sentencepiece) is used for word segmentation. The sentencepiece model is trained on CLUECorpusSmall corpus:
```
>>> import sentencepiece as spm
>>> spm.SentencePieceTrainer.train(input='cluecorpussmall.txt',
model_prefix='cluecorpussmall_spm',
vocab_size=100000,
max_sentence_length=1024,
max_sentencepiece_length=6,
user_defined_symbols=['[MASK]','[unused1]','[unused2]',
'[unused3]','[unused4]','[unused5]','[unused6]',
'[unused7]','[unused8]','[unused9]','[unused10]'],
pad_id=0,
pad_piece='[PAD]',
unk_id=1,
unk_piece='[UNK]',
bos_id=2,
bos_piece='[CLS]',
eos_id=3,
eos_piece='[SEP]',
train_extremely_large_corpus=True
)
```
## Training procedure
Models are pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 word-based RoBERTa-Medium
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq128_dataset.pt \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_word_roberta_medium_seq128_model.bin-1000000 \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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_word_roberta_medium_seq128_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 --target mlm
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/roberta-tiny-word-chinese-cluecorpussmall
[4_256]:https://huggingface.co/uer/roberta-mini-word-chinese-cluecorpussmall
[4_512]:https://huggingface.co/uer/roberta-small-word-chinese-cluecorpussmall
[8_512]:https://huggingface.co/uer/roberta-medium-word-chinese-cluecorpussmall
[12_768]:https://huggingface.co/uer/roberta-base-word-chinese-cluecorpussmall |
uer/roberta-medium-word-chinese-cluecorpussmall | 2021-05-20T08:43:30.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"zh",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tf_model.h5",
"tokenizer_config.json"
] | uer | 690 | transformers | ---
language: zh
datasets: CLUECorpusSmall
widget:
- text: "最近一趟去北京的[MASK]几点发车"
---
# Chinese word-based RoBERTa Miniatures
## Model description
This is the set of 5 Chinese word-based RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
Most Chinese pre-trained weights are based on Chinese character. Compared with character-based models, word-based models are faster (because of shorter sequence length) and have better performance according to our experimental results. To this end, we released the 5 Chinese word-based RoBERTa models of different sizes. In order to facilitate users to reproduce the results, we used the publicly available corpus and word segmentation tool, and provided all training details.
Notice that the output results of Hosted inference API (right) are not properly displayed. When the predicted word has multiple characters, the single word instead of entire sentence is displayed. One can click **JSON Output** for normal output results.
You can download the 5 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | Link |
| -------- | :-----------------------: |
| **word-based RoBERTa-Tiny** | [**L=2/H=128 (Tiny)**][2_128] |
| **word-based RoBERTa-Mini** | [**L=4/H=256 (Mini)**][4_256] |
| **word-based RoBERTa-Small** | [**L=4/H=512 (Small)**][4_512] |
| **word-based RoBERTa-Medium** | [**L=8/H=512 (Medium)**][8_512] |
| **word-based RoBERTa-Base** | [**L=12/H=768 (Base)**][12_768] |
Compared with [char-based models](https://huggingface.co/uer/chinese_roberta_L-2_H-128), word-based models achieve better results in most cases. Here are scores on the devlopment set of six Chinese tasks:
| Model | Score | douban | chnsenticorp | lcqmc | tnews(CLUE) | iflytek(CLUE) | ocnli(CLUE) |
| -------------- | :---: | :----: | :----------: | :---: | :---------: | :-----------: | :---------: |
| RoBERTa-Tiny(char) | 72.3 | 83.0 | 91.4 | 81.8 | 62.0 | 55.0 | 60.3 |
| **RoBERTa-Tiny(word)** | **74.3(+2.0)** | **86.4** | **93.2** | **82.0** | **66.4** | **58.2** | **59.6** |
| RoBERTa-Mini(char) | 75.7 | 84.8 | 93.7 | 86.1 | 63.9 | 58.3 | 67.4 |
| **RoBERTa-Mini(word)** | **76.7(+1.0)** | **87.6** | **94.1** | **85.4** | **66.9** | **59.2** | **67.3** |
| RoBERTa-Small(char) | 76.8 | 86.5 | 93.4 | 86.5 | 65.1 | 59.4 | 69.7 |
| **RoBERTa-Small(word)** | **78.1(+1.3)** | **88.5** | **94.7** | **87.4** | **67.6** | **60.9** | **69.8** |
| RoBERTa-Medium(char) | 77.8 | 87.6 | 94.8 | 88.1 | 65.6 | 59.5 | 71.2 |
| **RoBERTa-Medium(word)** | **78.9(+1.1)** | **89.2** | **95.1** | **88.0** | **67.8** | **60.6** | **73.0** |
| RoBERTa-Base(char) | 79.5 | 89.1 | 95.2 | 89.2 | 67.0 | 60.9 | 75.5 |
| **RoBERTa-Base(word)** | **80.2(+0.7)** | **90.3** | **95.7** | **89.4** | **68.0** | **61.5** | **76.8** |
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 word-based RoBERTa-Medium):
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-medium-word-chinese-cluecorpussmall')
>>> unmasker("[MASK]的首都是北京。")
[
{'sequence': '中国 的首都是北京。',
'score': 0.21525809168815613,
'token': 2873,
'token_str': '中国'},
{'sequence': '北京 的首都是北京。',
'score': 0.15194718539714813,
'token': 9502,
'token_str': '北京'},
{'sequence': '我们 的首都是北京。',
'score': 0.08854265511035919,
'token': 4215,
'token_str': '我们'},
{'sequence': '美国 的首都是北京。',
'score': 0.06808705627918243,
'token': 7810,
'token_str': '美国'},
{'sequence': '日本 的首都是北京。',
'score': 0.06071401759982109,
'token': 7788,
'token_str': '日本'}
]
```
Here is how to use this model to get the features of a given text in PyTorch:
```python
from transformers import AlbertTokenizer, BertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-medium-word-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:
```python
from transformers import AlbertTokenizer, TFBertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-medium-word-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)
```
Since BertTokenizer does not support sentencepiece, AlbertTokenizer is used here.
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data. Google's [sentencepiece](https://github.com/google/sentencepiece) is used for word segmentation. The sentencepiece model is trained on CLUECorpusSmall corpus:
```
>>> import sentencepiece as spm
>>> spm.SentencePieceTrainer.train(input='cluecorpussmall.txt',
model_prefix='cluecorpussmall_spm',
vocab_size=100000,
max_sentence_length=1024,
max_sentencepiece_length=6,
user_defined_symbols=['[MASK]','[unused1]','[unused2]',
'[unused3]','[unused4]','[unused5]','[unused6]',
'[unused7]','[unused8]','[unused9]','[unused10]'],
pad_id=0,
pad_piece='[PAD]',
unk_id=1,
unk_piece='[UNK]',
bos_id=2,
bos_piece='[CLS]',
eos_id=3,
eos_piece='[SEP]',
train_extremely_large_corpus=True
)
```
## Training procedure
Models are pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 word-based RoBERTa-Medium
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq128_dataset.pt \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_word_roberta_medium_seq128_model.bin-1000000 \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \\
--output_model_path models/cluecorpussmall_word_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_word_roberta_medium_seq128_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 --target mlm
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/roberta-tiny-word-chinese-cluecorpussmall
[4_256]:https://huggingface.co/uer/roberta-mini-word-chinese-cluecorpussmall
[4_512]:https://huggingface.co/uer/roberta-small-word-chinese-cluecorpussmall
[8_512]:https://huggingface.co/uer/roberta-medium-word-chinese-cluecorpussmall
[12_768]:https://huggingface.co/uer/roberta-base-word-chinese-cluecorpussmall |
uer/roberta-mini-word-chinese-cluecorpussmall | 2021-05-20T08:44:09.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"zh",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tf_model.h5",
"tokenizer_config.json"
] | uer | 27 | transformers | ---
language: zh
datasets: CLUECorpusSmall
widget:
- text: "最近一趟去北京的[MASK]几点发车"
---
# Chinese word-based RoBERTa Miniatures
## Model description
This is the set of 5 Chinese word-based RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
Most Chinese pre-trained weights are based on Chinese character. Compared with character-based models, word-based models are faster (because of shorter sequence length) and have better performance according to our experimental results. To this end, we released the 5 Chinese word-based RoBERTa models of different sizes. In order to facilitate users to reproduce the results, we used the publicly available corpus and word segmentation tool, and provided all training details.
Notice that the output results of Hosted inference API (right) are not properly displayed. When the predicted word has multiple characters, the single word instead of entire sentence is displayed. One can click **JSON Output** for normal output results.
You can download the 5 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | Link |
| -------- | :-----------------------: |
| **word-based RoBERTa-Tiny** | [**L=2/H=128 (Tiny)**][2_128] |
| **word-based RoBERTa-Mini** | [**L=4/H=256 (Mini)**][4_256] |
| **word-based RoBERTa-Small** | [**L=4/H=512 (Small)**][4_512] |
| **word-based RoBERTa-Medium** | [**L=8/H=512 (Medium)**][8_512] |
| **word-based RoBERTa-Base** | [**L=12/H=768 (Base)**][12_768] |
Compared with [char-based models](https://huggingface.co/uer/chinese_roberta_L-2_H-128), word-based models achieve better results in most cases. Here are scores on the devlopment set of six Chinese tasks:
| Model | Score | douban | chnsenticorp | lcqmc | tnews(CLUE) | iflytek(CLUE) | ocnli(CLUE) |
| -------------- | :---: | :----: | :----------: | :---: | :---------: | :-----------: | :---------: |
| RoBERTa-Tiny(char) | 72.3 | 83.0 | 91.4 | 81.8 | 62.0 | 55.0 | 60.3 |
| **RoBERTa-Tiny(word)** | **74.3(+2.0)** | **86.4** | **93.2** | **82.0** | **66.4** | **58.2** | **59.6** |
| RoBERTa-Mini(char) | 75.7 | 84.8 | 93.7 | 86.1 | 63.9 | 58.3 | 67.4 |
| **RoBERTa-Mini(word)** | **76.7(+1.0)** | **87.6** | **94.1** | **85.4** | **66.9** | **59.2** | **67.3** |
| RoBERTa-Small(char) | 76.8 | 86.5 | 93.4 | 86.5 | 65.1 | 59.4 | 69.7 |
| **RoBERTa-Small(word)** | **78.1(+1.3)** | **88.5** | **94.7** | **87.4** | **67.6** | **60.9** | **69.8** |
| RoBERTa-Medium(char) | 77.8 | 87.6 | 94.8 | 88.1 | 65.6 | 59.5 | 71.2 |
| **RoBERTa-Medium(word)** | **78.9(+1.1)** | **89.2** | **95.1** | **88.0** | **67.8** | **60.6** | **73.0** |
| RoBERTa-Base(char) | 79.5 | 89.1 | 95.2 | 89.2 | 67.0 | 60.9 | 75.5 |
| **RoBERTa-Base(word)** | **80.2(+0.7)** | **90.3** | **95.7** | **89.4** | **68.0** | **61.5** | **76.8** |
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 word-based RoBERTa-Medium):
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-medium-word-chinese-cluecorpussmall')
>>> unmasker("[MASK]的首都是北京。")
[
{'sequence': '中国 的首都是北京。',
'score': 0.21525809168815613,
'token': 2873,
'token_str': '中国'},
{'sequence': '北京 的首都是北京。',
'score': 0.15194718539714813,
'token': 9502,
'token_str': '北京'},
{'sequence': '我们 的首都是北京。',
'score': 0.08854265511035919,
'token': 4215,
'token_str': '我们'},
{'sequence': '美国 的首都是北京。',
'score': 0.06808705627918243,
'token': 7810,
'token_str': '美国'},
{'sequence': '日本 的首都是北京。',
'score': 0.06071401759982109,
'token': 7788,
'token_str': '日本'}
]
```
Here is how to use this model to get the features of a given text in PyTorch:
```python
from transformers import AlbertTokenizer, BertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-medium-word-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:
```python
from transformers import AlbertTokenizer, TFBertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-medium-word-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)
```
Since BertTokenizer does not support sentencepiece, AlbertTokenizer is used here.
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data. Google's [sentencepiece](https://github.com/google/sentencepiece) is used for word segmentation. The sentencepiece model is trained on CLUECorpusSmall corpus:
```
>>> import sentencepiece as spm
>>> spm.SentencePieceTrainer.train(input='cluecorpussmall.txt',
model_prefix='cluecorpussmall_spm',
vocab_size=100000,
max_sentence_length=1024,
max_sentencepiece_length=6,
user_defined_symbols=['[MASK]','[unused1]','[unused2]',
'[unused3]','[unused4]','[unused5]','[unused6]',
'[unused7]','[unused8]','[unused9]','[unused10]'],
pad_id=0,
pad_piece='[PAD]',
unk_id=1,
unk_piece='[UNK]',
bos_id=2,
bos_piece='[CLS]',
eos_id=3,
eos_piece='[SEP]',
train_extremely_large_corpus=True
)
```
## Training procedure
Models are pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 word-based RoBERTa-Medium
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq128_dataset.pt \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_word_roberta_medium_seq128_model.bin-1000000 \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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_word_roberta_medium_seq128_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 --target mlm
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/roberta-tiny-word-chinese-cluecorpussmall
[4_256]:https://huggingface.co/uer/roberta-mini-word-chinese-cluecorpussmall
[4_512]:https://huggingface.co/uer/roberta-small-word-chinese-cluecorpussmall
[8_512]:https://huggingface.co/uer/roberta-medium-word-chinese-cluecorpussmall
[12_768]:https://huggingface.co/uer/roberta-base-word-chinese-cluecorpussmall |
uer/roberta-small-word-chinese-cluecorpussmall | 2021-05-20T08:44:51.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"zh",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tf_model.h5",
"tokenizer_config.json"
] | uer | 325 | transformers | ---
language: zh
datasets: CLUECorpusSmall
widget:
- text: "最近一趟去北京的[MASK]几点发车"
---
# Chinese word-based RoBERTa Miniatures
## Model description
This is the set of 5 Chinese word-based RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
Most Chinese pre-trained weights are based on Chinese character. Compared with character-based models, word-based models are faster (because of shorter sequence length) and have better performance according to our experimental results. To this end, we released the 5 Chinese word-based RoBERTa models of different sizes. In order to facilitate users to reproduce the results, we used the publicly available corpus and word segmentation tool, and provided all training details.
Notice that the output results of Hosted inference API (right) are not properly displayed. When the predicted word has multiple characters, the single word instead of entire sentence is displayed. One can click **JSON Output** for normal output results.
You can download the 5 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | Link |
| -------- | :-----------------------: |
| **word-based RoBERTa-Tiny** | [**L=2/H=128 (Tiny)**][2_128] |
| **word-based RoBERTa-Mini** | [**L=4/H=256 (Mini)**][4_256] |
| **word-based RoBERTa-Small** | [**L=4/H=512 (Small)**][4_512] |
| **word-based RoBERTa-Medium** | [**L=8/H=512 (Medium)**][8_512] |
| **word-based RoBERTa-Base** | [**L=12/H=768 (Base)**][12_768] |
Compared with [char-based models](https://huggingface.co/uer/chinese_roberta_L-2_H-128), word-based models achieve better results in most cases. Here are scores on the devlopment set of six Chinese tasks:
| Model | Score | douban | chnsenticorp | lcqmc | tnews(CLUE) | iflytek(CLUE) | ocnli(CLUE) |
| -------------- | :---: | :----: | :----------: | :---: | :---------: | :-----------: | :---------: |
| RoBERTa-Tiny(char) | 72.3 | 83.0 | 91.4 | 81.8 | 62.0 | 55.0 | 60.3 |
| **RoBERTa-Tiny(word)** | **74.3(+2.0)** | **86.4** | **93.2** | **82.0** | **66.4** | **58.2** | **59.6** |
| RoBERTa-Mini(char) | 75.7 | 84.8 | 93.7 | 86.1 | 63.9 | 58.3 | 67.4 |
| **RoBERTa-Mini(word)** | **76.7(+1.0)** | **87.6** | **94.1** | **85.4** | **66.9** | **59.2** | **67.3** |
| RoBERTa-Small(char) | 76.8 | 86.5 | 93.4 | 86.5 | 65.1 | 59.4 | 69.7 |
| **RoBERTa-Small(word)** | **78.1(+1.3)** | **88.5** | **94.7** | **87.4** | **67.6** | **60.9** | **69.8** |
| RoBERTa-Medium(char) | 77.8 | 87.6 | 94.8 | 88.1 | 65.6 | 59.5 | 71.2 |
| **RoBERTa-Medium(word)** | **78.9(+1.1)** | **89.2** | **95.1** | **88.0** | **67.8** | **60.6** | **73.0** |
| RoBERTa-Base(char) | 79.5 | 89.1 | 95.2 | 89.2 | 67.0 | 60.9 | 75.5 |
| **RoBERTa-Base(word)** | **80.2(+0.7)** | **90.3** | **95.7** | **89.4** | **68.0** | **61.5** | **76.8** |
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 word-based RoBERTa-Medium):
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-medium-word-chinese-cluecorpussmall')
>>> unmasker("[MASK]的首都是北京。")
[
{'sequence': '中国 的首都是北京。',
'score': 0.21525809168815613,
'token': 2873,
'token_str': '中国'},
{'sequence': '北京 的首都是北京。',
'score': 0.15194718539714813,
'token': 9502,
'token_str': '北京'},
{'sequence': '我们 的首都是北京。',
'score': 0.08854265511035919,
'token': 4215,
'token_str': '我们'},
{'sequence': '美国 的首都是北京。',
'score': 0.06808705627918243,
'token': 7810,
'token_str': '美国'},
{'sequence': '日本 的首都是北京。',
'score': 0.06071401759982109,
'token': 7788,
'token_str': '日本'}
]
```
Here is how to use this model to get the features of a given text in PyTorch:
```python
from transformers import AlbertTokenizer, BertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-medium-word-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:
```python
from transformers import AlbertTokenizer, TFBertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-medium-word-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)
```
Since BertTokenizer does not support sentencepiece, AlbertTokenizer is used here.
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data. Google's [sentencepiece](https://github.com/google/sentencepiece) is used for word segmentation. The sentencepiece model is trained on CLUECorpusSmall corpus:
```
>>> import sentencepiece as spm
>>> spm.SentencePieceTrainer.train(input='cluecorpussmall.txt',
model_prefix='cluecorpussmall_spm',
vocab_size=100000,
max_sentence_length=1024,
max_sentencepiece_length=6,
user_defined_symbols=['[MASK]','[unused1]','[unused2]',
'[unused3]','[unused4]','[unused5]','[unused6]',
'[unused7]','[unused8]','[unused9]','[unused10]'],
pad_id=0,
pad_piece='[PAD]',
unk_id=1,
unk_piece='[UNK]',
bos_id=2,
bos_piece='[CLS]',
eos_id=3,
eos_piece='[SEP]',
train_extremely_large_corpus=True
)
```
## Training procedure
Models are pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 word-based RoBERTa-Medium
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq128_dataset.pt \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_word_roberta_medium_seq128_model.bin-1000000 \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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_word_roberta_medium_seq128_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 --target mlm
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/roberta-tiny-word-chinese-cluecorpussmall
[4_256]:https://huggingface.co/uer/roberta-mini-word-chinese-cluecorpussmall
[4_512]:https://huggingface.co/uer/roberta-small-word-chinese-cluecorpussmall
[8_512]:https://huggingface.co/uer/roberta-medium-word-chinese-cluecorpussmall
[12_768]:https://huggingface.co/uer/roberta-base-word-chinese-cluecorpussmall |
uer/roberta-tiny-word-chinese-cluecorpussmall | 2021-05-20T08:46:02.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"masked-lm",
"zh",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tf_model.h5",
"tokenizer_config.json"
] | uer | 280 | transformers | ---
language: zh
datasets: CLUECorpusSmall
widget:
- text: "最近一趟去北京的[MASK]几点发车"
---
# Chinese word-based RoBERTa Miniatures
## Model description
This is the set of 5 Chinese word-based RoBERTa models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
Most Chinese pre-trained weights are based on Chinese character. Compared with character-based models, word-based models are faster (because of shorter sequence length) and have better performance according to our experimental results. To this end, we released the 5 Chinese word-based RoBERTa models of different sizes. In order to facilitate users to reproduce the results, we used the publicly available corpus and word segmentation tool, and provided all training details.
Notice that the output results of Hosted inference API (right) are not properly displayed. When the predicted word has multiple characters, the single word instead of entire sentence is displayed. One can click **JSON Output** for normal output results.
You can download the 5 Chinese RoBERTa miniatures either from the [UER-py Github page](https://github.com/dbiir/UER-py/), or via HuggingFace from the links below:
| | Link |
| -------- | :-----------------------: |
| **word-based RoBERTa-Tiny** | [**L=2/H=128 (Tiny)**][2_128] |
| **word-based RoBERTa-Mini** | [**L=4/H=256 (Mini)**][4_256] |
| **word-based RoBERTa-Small** | [**L=4/H=512 (Small)**][4_512] |
| **word-based RoBERTa-Medium** | [**L=8/H=512 (Medium)**][8_512] |
| **word-based RoBERTa-Base** | [**L=12/H=768 (Base)**][12_768] |
Compared with [char-based models](https://huggingface.co/uer/chinese_roberta_L-2_H-128), word-based models achieve better results in most cases. Here are scores on the devlopment set of six Chinese tasks:
| Model | Score | douban | chnsenticorp | lcqmc | tnews(CLUE) | iflytek(CLUE) | ocnli(CLUE) |
| -------------- | :---: | :----: | :----------: | :---: | :---------: | :-----------: | :---------: |
| RoBERTa-Tiny(char) | 72.3 | 83.0 | 91.4 | 81.8 | 62.0 | 55.0 | 60.3 |
| **RoBERTa-Tiny(word)** | **74.3(+2.0)** | **86.4** | **93.2** | **82.0** | **66.4** | **58.2** | **59.6** |
| RoBERTa-Mini(char) | 75.7 | 84.8 | 93.7 | 86.1 | 63.9 | 58.3 | 67.4 |
| **RoBERTa-Mini(word)** | **76.7(+1.0)** | **87.6** | **94.1** | **85.4** | **66.9** | **59.2** | **67.3** |
| RoBERTa-Small(char) | 76.8 | 86.5 | 93.4 | 86.5 | 65.1 | 59.4 | 69.7 |
| **RoBERTa-Small(word)** | **78.1(+1.3)** | **88.5** | **94.7** | **87.4** | **67.6** | **60.9** | **69.8** |
| RoBERTa-Medium(char) | 77.8 | 87.6 | 94.8 | 88.1 | 65.6 | 59.5 | 71.2 |
| **RoBERTa-Medium(word)** | **78.9(+1.1)** | **89.2** | **95.1** | **88.0** | **67.8** | **60.6** | **73.0** |
| RoBERTa-Base(char) | 79.5 | 89.1 | 95.2 | 89.2 | 67.0 | 60.9 | 75.5 |
| **RoBERTa-Base(word)** | **80.2(+0.7)** | **90.3** | **95.7** | **89.4** | **68.0** | **61.5** | **76.8** |
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 word-based RoBERTa-Medium):
```python
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='uer/roberta-medium-word-chinese-cluecorpussmall')
>>> unmasker("[MASK]的首都是北京。")
[
{'sequence': '中国 的首都是北京。',
'score': 0.21525809168815613,
'token': 2873,
'token_str': '中国'},
{'sequence': '北京 的首都是北京。',
'score': 0.15194718539714813,
'token': 9502,
'token_str': '北京'},
{'sequence': '我们 的首都是北京。',
'score': 0.08854265511035919,
'token': 4215,
'token_str': '我们'},
{'sequence': '美国 的首都是北京。',
'score': 0.06808705627918243,
'token': 7810,
'token_str': '美国'},
{'sequence': '日本 的首都是北京。',
'score': 0.06071401759982109,
'token': 7788,
'token_str': '日本'}
]
```
Here is how to use this model to get the features of a given text in PyTorch:
```python
from transformers import AlbertTokenizer, BertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = BertModel.from_pretrained("uer/roberta-medium-word-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:
```python
from transformers import AlbertTokenizer, TFBertModel
tokenizer = AlbertTokenizer.from_pretrained('uer/roberta-medium-word-chinese-cluecorpussmall')
model = TFBertModel.from_pretrained("uer/roberta-medium-word-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)
```
Since BertTokenizer does not support sentencepiece, AlbertTokenizer is used here.
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data. Google's [sentencepiece](https://github.com/google/sentencepiece) is used for word segmentation. The sentencepiece model is trained on CLUECorpusSmall corpus:
```
>>> import sentencepiece as spm
>>> spm.SentencePieceTrainer.train(input='cluecorpussmall.txt',
model_prefix='cluecorpussmall_spm',
vocab_size=100000,
max_sentence_length=1024,
max_sentencepiece_length=6,
user_defined_symbols=['[MASK]','[unused1]','[unused2]',
'[unused3]','[unused4]','[unused5]','[unused6]',
'[unused7]','[unused8]','[unused9]','[unused10]'],
pad_id=0,
pad_piece='[PAD]',
unk_id=1,
unk_piece='[UNK]',
bos_id=2,
bos_piece='[CLS]',
eos_id=3,
eos_piece='[SEP]',
train_extremely_large_corpus=True
)
```
## Training procedure
Models are pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 word-based RoBERTa-Medium
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq128_dataset.pt \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--spm_model_path models/cluecorpussmall_spm.model \
--dataset_path cluecorpussmall_word_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target mlm
```
```
python3 pretrain.py --dataset_path cluecorpussmall_word_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_word_roberta_medium_seq128_model.bin-1000000 \
--spm_model_path models/cluecorpussmall_spm.model \
--config_path models/bert/medium_config.json \
--output_model_path models/cluecorpussmall_word_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_word_roberta_medium_seq128_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 --target mlm
```
### BibTeX entry and citation info
```
@article{devlin2018bert,
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},
year={2018}
}
@article{turc2019,
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 },
year={2019}
}
@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}
}
```
[2_128]:https://huggingface.co/uer/roberta-tiny-word-chinese-cluecorpussmall
[4_256]:https://huggingface.co/uer/roberta-mini-word-chinese-cluecorpussmall
[4_512]:https://huggingface.co/uer/roberta-small-word-chinese-cluecorpussmall
[8_512]:https://huggingface.co/uer/roberta-medium-word-chinese-cluecorpussmall
[12_768]:https://huggingface.co/uer/roberta-base-word-chinese-cluecorpussmall |
uer/t5-base-chinese-cluecorpussmall | 2021-05-28T09:22:47.000Z | [
"pytorch",
"tf",
"t5",
"seq2seq",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 521 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "作为电子extra0的平台，京东绝对是领先者。如今的刘强extra1已经是身价过extra2的老板。"
---
# Chinese T5
## Model description
This is the set of Chinese T5 models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
The Text-to-Text Transfer Transformer (T5) leverages a unified text-to-text format and attains state-of-the-art results on a wide variety of English-language NLP tasks. Following their work, we released a series of Chinese T5 models.
| | Link |
| -------- | :-----------------------: |
| **T5-Small** | [**L=6/H=512 (Small)**][small] |
| **T5-Base** | [**L=12/H=768 (Base)**][base] |
In T5, spans of the input sequence are masked by so-called sentinel token. Each sentinel token represents a unique mask token for the input sequence and should start with `<extra_id_0>`, `<extra_id_1>`, … up to `<extra_id_99>`. However, `<extra_id_xxx>` is separated into multiple parts in Huggingface's Hosted inference API. Therefore, we replace `<extra_id_xxx>` with `extraxxx` in vocabulary and BertTokenizer regards `extraxxx` as one sentinel token.
## How to use
You can use this model directly with a pipeline for text2text generation (take the case of T5-Small):
```python
>>> from transformers import BertTokenizer, T5ForConditionalGeneration, Text2TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/t5-small-chinese-cluecorpussmall")
>>> model = T5ForConditionalGeneration.from_pretrained("uer/t5-small-chinese-cluecorpussmall")
>>> text2text_generator = Text2TextGenerationPipeline(model, tokenizer)
>>> text2text_generator("中国的首都是extra0京", max_length=50, do_sample=False)
[{'generated_text': 'extra0 北 extra1 extra2 extra3 extra4 extra5'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 T5-Small
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5_seq128_dataset.pt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5/small_config.json \
--output_model_path models/cluecorpussmall_t5_small_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-3 --batch_size 64 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre --decoder transformer \
--target t5 --tie_weights
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5_small_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_t5_small_seq128_model.bin-1000000 \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5/small_config.json \
--output_model_path models/cluecorpussmall_t5_small_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-4 --batch_size 16 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre --decoder transformer \
--target t5 --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_t5_from_uer_to_huggingface.py --input_model_path cluecorpussmall_t5_small_seq512_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 6 \
--type t5
```
### BibTeX entry and citation info
```
@article{2020t5,
title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
journal = {Journal of Machine Learning Research},
pages = {1-67},
year = {2020}
}
@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}
}
```
[small]:https://huggingface.co/uer/t5-small-chinese-cluecorpussmall
[base]:https://huggingface.co/uer/t5-base-chinese-cluecorpussmall |
uer/t5-small-chinese-cluecorpussmall | 2021-04-04T03:41:04.000Z | [
"pytorch",
"tf",
"t5",
"seq2seq",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 608 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "作为电子extra0的平台，京东绝对是领先者。如今的刘强extra1已经是身价过extra2的老板。"
---
# Chinese T5
## Model description
This is the set of Chinese T5 models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
The Text-to-Text Transfer Transformer (T5) leverages a unified text-to-text format and attains state-of-the-art results on a wide variety of English-language NLP tasks. Following their work, we released a series of Chinese T5 models.
| | Link |
| -------- | :-----------------------: |
| **T5-Small** | [**L=6/H=512 (Small)**][small] |
| **T5-Base** | [**L=12/H=768 (Base)**][base] |
In T5, spans of the input sequence are masked by so-called sentinel token. Each sentinel token represents a unique mask token for the input sequence and should start with `<extra_id_0>`, `<extra_id_1>`, … up to `<extra_id_99>`. However, `<extra_id_xxx>` is separated into multiple parts in Huggingface's Hosted inference API. Therefore, we replace `<extra_id_xxx>` with `extraxxx` in vocabulary and BertTokenizer regards `extraxxx` as one sentinel token.
## How to use
You can use this model directly with a pipeline for text2text generation (take the case of T5-Small):
```python
>>> from transformers import BertTokenizer, T5ForConditionalGeneration, Text2TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/t5-small-chinese-cluecorpussmall")
>>> model = T5ForConditionalGeneration.from_pretrained("uer/t5-small-chinese-cluecorpussmall")
>>> text2text_generator = Text2TextGenerationPipeline(model, tokenizer)
>>> text2text_generator("中国的首都是extra0京", max_length=50, do_sample=False)
[{'generated_text': 'extra0 北 extra1 extra2 extra3 extra4 extra5'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 T5-Small
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5_seq128_dataset.pt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5/small_config.json \
--output_model_path models/cluecorpussmall_t5_small_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-3 --batch_size 64 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre --decoder transformer \
--target t5 --tie_weights
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5_small_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_t5_small_seq128_model.bin-1000000 \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5/small_config.json \
--output_model_path models/cluecorpussmall_t5_small_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-4 --batch_size 16 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre --decoder transformer \
--target t5 --tie_weights
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_t5_from_uer_to_huggingface.py --input_model_path cluecorpussmall_t5_small_seq512_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 6 \
--type t5
```
### BibTeX entry and citation info
```
@article{2020t5,
title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
journal = {Journal of Machine Learning Research},
pages = {1-67},
year = {2020}
}
@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}
}
```
[small]:https://huggingface.co/uer/t5-small-chinese-cluecorpussmall
[base]:https://huggingface.co/uer/t5-base-chinese-cluecorpussmall |
uer/t5-v1_1-small-chinese-cluecorpussmall | 2021-04-05T13:58:15.000Z | [
"pytorch",
"tf",
"mt5",
"seq2seq",
"Chinese",
"dataset:CLUECorpusSmall",
"arxiv:1909.05658",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | uer | 177 | transformers | ---
language: Chinese
datasets: CLUECorpusSmall
widget:
- text: "作为电子extra0的平台，京东绝对是领先者。如今的刘强extra1已经是身价过extra2的老板。"
---
# Chinese T5 Version 1.1
## Model description
This is the set of Chinese T5 Version 1.1 models pre-trained by [UER-py](https://arxiv.org/abs/1909.05658).
**Version 1.1**
Chinese T5 Version 1.1 includes the following improvements compared to our Chinese T5 model:
- GEGLU activation in feed-forward hidden layer, rather than ReLU
- Dropout was turned off in pre-training
- no parameter sharing between embedding and classifier layer
| | Link |
| ----------------- | :----------------------------: |
| **T5-v1_1-Small** | [**L=8/H=512 (Small)**][small] |
| **T5-v1_1-Base** | [**L=12/H=768 (Base)**][base] |
In T5 Version 1.1, spans of the input sequence are masked by so-called sentinel token. Each sentinel token represents a unique mask token for the input sequence and should start with `<extra_id_0>`, `<extra_id_1>`, … up to `<extra_id_99>`. However, `<extra_id_xxx>` is separated into multiple parts in Huggingface's Hosted inference API. Therefore, we replace `<extra_id_xxx>` with `extraxxx` in vocabulary and BertTokenizer regards `extraxxx` as one sentinel token.
## How to use
You can use this model directly with a pipeline for text2text generation (take the case of T5-v1_1-Small):
```python
>>> from transformers import BertTokenizer, MT5ForConditionalGeneration, Text2TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/t5-v1_1-small-chinese-cluecorpussmall")
>>> model = MT5ForConditionalGeneration.from_pretrained("uer/t5-v1_1-small-chinese-cluecorpussmall")
>>> text2text_generator = Text2TextGenerationPipeline(model, tokenizer)
>>> text2text_generator("中国的首都是extra0京", max_length=50, do_sample=False)
[{'generated_text': 'extra0 北 extra1 extra2 extra3 extra4 extra5'}]
```
## Training data
[CLUECorpusSmall](https://github.com/CLUEbenchmark/CLUECorpus2020/) is used as training data.
## Training procedure
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). 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 T5-v1_1-Small
Stage1:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5-v1_1_seq128_dataset.pt \
--processes_num 32 --seq_length 128 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5-v1_1_seq128_dataset.pt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5-v1_1/small_config.json \
--output_model_path models/cluecorpussmall_t5-v1_1_small_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-3 --batch_size 64 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre \
--feed_forward gated --decoder transformer --target t5
```
Stage2:
```
python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--dataset_path cluecorpussmall_t5-v1_1_seq512_dataset.pt \
--processes_num 32 --seq_length 512 \
--dynamic_masking --target t5
```
```
python3 pretrain.py --dataset_path cluecorpussmall_t5-v1_1_seq512_dataset.pt \
--pretrained_model_path models/cluecorpussmall_t5-v1_1_small_seq128_model.bin-1000000 \
--vocab_path models/google_zh_with_sentinel_vocab.txt \
--config_path models/t5-v1_1/small_config.json \
--output_model_path models/cluecorpussmall_t5-v1_1_small_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-4 --batch_size 16 \
--span_masking --span_geo_prob 0.3 --span_max_length 5 \
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
--encoder transformer --mask fully_visible --layernorm_positioning pre \
--feed_forward gated --decoder transformer --target t5
```
Finally, we convert the pre-trained model into Huggingface's format:
```
python3 scripts/convert_t5_from_uer_to_huggingface.py --input_model_path cluecorpussmall_t5_small_seq512_model.bin-250000 \
--output_model_path pytorch_model.bin \
--layers_num 8 \
--type t5-v1_1
```
### BibTeX entry and citation info
```
@article{2020t5,
title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
journal = {Journal of Machine Learning Research},
pages = {1-67},
year = {2020}
}
@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}
}
```
[small]:https://huggingface.co/uer/t5-v1_1-small-chinese-cluecorpussmall
[base]:https://huggingface.co/uer/t5-v1_1-base-chinese-cluecorpussmall |
ufal/robeczech-base | 2021-05-30T08:00:34.000Z | [
"pytorch",
"tf",
"roberta",
"masked-lm",
"cs",
"arxiv:2105.11314",
"transformers",
"Czech",
"RoBERTa",
"ÚFAL",
"license:cc-by-nc-4.0",
"fill-mask"
] | fill-mask | [
".gitattributes",
"LICENSE",
"README.md",
"config.json",
"merges.txt",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer_config.json",
"vocab.json"
] | ufal | 467 | transformers | ---
language: "cs"
tags:
- Czech
- RoBERTa
- ÚFAL
license: "cc-by-nc-4.0"
---
# RobeCzech model
RobeCzech is a monolingual RoBERTa language representation model trained on Czech data.
RobeCzech model is released publicly at [LINDAT](https://hdl.handle.net/11234/1-3691) and [Hugging Face](https://huggingface.co/ufal/robeczech-base).
Please cite the corresponding publication:
- Milan Straka, Jakub Náplava, Jana Straková and David Samuel: Czech RoBERTa, a monolingual contextualized language representation model. Accepted to TSD 2021.
Preprint of the paper is available at https://arxiv.org/abs/2105.11314.
|
ugaray96/biobert_ncbi_disease_ner | 2021-05-20T08:46:47.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"token-classification",
"transformers"
] | token-classification | [
".gitattributes",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | ugaray96 | 566 | transformers | |
uklfr/gottbert-base | 2021-05-20T22:50:19.000Z | [
"pytorch",
"jax",
"roberta",
"masked-lm",
"arxiv:2012.02110",
"transformers",
"fill-mask"
] | fill-mask | [
".gitattributes",
"LICENSE",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"vocab.json"
] | uklfr | 1,680 | transformers | # Gottbert-base
BERT model trained solely on the German portion of the OSCAR data set.
[Paper: GottBERT: a pure German Language Model](https://arxiv.org/abs/2012.02110)
Authors: Raphael Scheible, Fabian Thomczyk, Patric Tippmann, Victor Jaravine, Martin Boeker |
unc-nlp/lxmert-base-uncased | 2021-03-10T02:39:25.000Z | [
"pytorch",
"tf",
"lxmert",
"transformers"
] | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
] | unc-nlp | 4,584 | transformers | ||
unc-nlp/lxmert-gqa-uncased | 2020-09-08T19:05:59.000Z | [
"pytorch",
"lxmert",
"question-answering",
"transformers"
] | question-answering | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.txt"
] | unc-nlp | 310 | transformers | |
unc-nlp/lxmert-vqa-uncased | 2020-09-10T17:57:42.000Z | [
"pytorch",
"lxmert",
"question-answering",
"transformers"
] | question-answering | [
".gitattributes",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"vocab.txt"
] | unc-nlp | 413 | transformers | |
unicamp-dl/ptt5-base-portuguese-vocab | 2021-03-24T22:16:54.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 334 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unicamp-dl/ptt5-base-t5-vocab | 2021-03-24T22:17:23.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 52 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unicamp-dl/ptt5-large-portuguese-vocab | 2021-03-24T22:17:55.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 77 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unicamp-dl/ptt5-large-t5-vocab | 2021-03-24T22:30:20.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 45 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unicamp-dl/ptt5-small-portuguese-vocab | 2021-03-24T22:18:14.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 159 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unicamp-dl/ptt5-small-t5-vocab | 2021-03-24T22:15:10.000Z | [
"pytorch",
"tf",
"t5",
"lm-head",
"seq2seq",
"pt",
"dataset:brWaC",
"transformers",
"license:mit",
"tensorflow",
"pt-br",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.h5"
] | unicamp-dl | 30 | transformers | ---
language: pt
license: mit
tags:
- t5
- pytorch
- tensorflow
- pt
- pt-br
datasets:
- brWaC
widget:
- text: "Texto de exemplo em português"
inference: false
---
# Portuguese T5 (aka "PTT5")
## Introduction
PTT5 is a T5 model pretrained in the BrWac corpus, a large collection of web pages in Portuguese, improving T5's performance on Portuguese sentence similarity and entailment tasks. It's available in three sizes (small, base and large) and two vocabularies (Google's T5 original and ours, trained on Portuguese Wikipedia).
For further information or requests, please go to [PTT5 repository](https://github.com/unicamp-dl/PTT5).
## Available models
| Model | Size | #Params | Vocabulary |
| :-: | :-: | :-: | :-: |
| [unicamp-dl/ptt5-small-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-small-t5-vocab) | small | 60M | Google's T5 |
| [unicamp-dl/ptt5-base-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-base-t5-vocab) | base | 220M | Google's T5 |
| [unicamp-dl/ptt5-large-t5-vocab](https://huggingface.co/unicamp-dl/ptt5-large-t5-vocab) | large | 740M | Google's T5 |
| [unicamp-dl/ptt5-small-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-small-portuguese-vocab) | small | 60M | Portuguese |
| **[unicamp-dl/ptt5-base-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-base-portuguese-vocab)** **(Recommended)** | **base** | **220M** | **Portuguese** |
| [unicamp-dl/ptt5-large-portuguese-vocab](https://huggingface.co/unicamp-dl/ptt5-large-portuguese-vocab) | large | 740M | Portuguese |
## Usage
```python
# Tokenizer
from transformers import T5Tokenizer
# PyTorch (bare model, baremodel + language modeling head)
from transformers import T5Model, T5ForConditionalGeneration
# Tensorflow (bare model, baremodel + language modeling head)
from transformers import TFT5Model, TFT5ForConditionalGeneration
model_name = 'unicamp-dl/ptt5-base-portuguese-vocab'
tokenizer = T5Tokenizer.from_pretrained(model_name)
# PyTorch
model_pt = T5ForConditionalGeneration.from_pretrained(model_name)
# TensorFlow
model_tf = TFT5ForConditionalGeneration.from_pretrained(model_name)
```
# Citation
If you use PTT5, please cite:
@article{ptt5_2020,
title={PTT5: Pretraining and validating the T5 model on Brazilian Portuguese data},
author={Carmo, Diedre and Piau, Marcos and Campiotti, Israel and Nogueira, Rodrigo and Lotufo, Roberto},
journal={arXiv preprint arXiv:2008.09144},
year={2020}
}
|
unideeplearning/polibert_sa | 2021-05-20T08:47:45.000Z | [
"pytorch",
"tf",
"jax",
"bert",
"text-classification",
"it",
"transformers",
"sentiment",
"Italian",
"license:mit"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.txt"
] | unideeplearning | 84 | transformers | ---
language: it
tags:
- sentiment
- Italian
license: MIT
widget:
- text: 'Giuseppe Rossi è un ottimo politico'
---
# 🤗 + polibert_SA - POLItic BERT based Sentiment Analysis
## Model description
This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets. You can try it out at https://www.unideeplearning.com/twitter_sa/ (in italian!)
#### Hands-on
```python
import torch
from torch import nn
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("unideeplearning/polibert_sa")
model = AutoModelForSequenceClassification.from_pretrained("unideeplearning/polibert_sa")
text = "Giuseppe Rossi è un pessimo politico"
input_ids = tokenizer.encode(text, add_special_tokens=True, return_tensors= 'pt')
logits, = model(input_ids)
logits = logits.squeeze(0)
prob = nn.functional.softmax(logits, dim=0)
# 0 Negative, 1 Neutral, 2 Positive
print(prob.argmax().tolist())
```
#### Hyperparameters
- Optimizer: **AdamW** with learning rate of **2e-5**, epsilon of **1e-8**
- Max epochs: **2**
- Batch size: **16**
## Acknowledgments
Thanks to the support from:
the [Hugging Face](https://huggingface.co/), https://www.unioneprofessionisti.com
https://www.unideeplearning.com/
|
unitary/multilingual-toxic-xlm-roberta | 2021-05-06T11:04:34.000Z | [
"pytorch",
"xlm-roberta",
"text-classification",
"arxiv:1703.04009",
"arxiv:1905.12516",
"transformers",
"pipeline_tag:text-classification"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json"
] | unitary | 197 | transformers | ---
pipeline_tag: "text-classification"
---
<div align="center">
**⚠️ Disclaimer:**
The huggingface models currently give different results to the detoxify library (see issue [here](https://github.com/unitaryai/detoxify/issues/15)). For the most up to date models we recommend using the models from https://github.com/unitaryai/detoxify
# 🙊 Detoxify
## Toxic Comment Classification with ⚡ Pytorch Lightning and 🤗 Transformers
</div>
## Description
Trained models & code to predict toxic comments on 3 Jigsaw challenges: Toxic comment classification, Unintended Bias in Toxic comments, Multilingual toxic comment classification.
Built by [Laura Hanu](https://laurahanu.github.io/) at [Unitary](https://www.unitary.ai/), where we are working to stop harmful content online by interpreting visual content in context.
Dependencies:
- For inference:
- 🤗 Transformers
- ⚡ Pytorch lightning
- For training will also need:
- Kaggle API (to download data)
| Challenge | Year | Goal | Original Data Source | Detoxify Model Name | Top Kaggle Leaderboard Score | Detoxify Score
|-|-|-|-|-|-|-|
| [Toxic Comment Classification Challenge](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge) | 2018 | build a multi-headed model that’s capable of detecting different types of of toxicity like threats, obscenity, insults, and identity-based hate. | Wikipedia Comments | `original` | 0.98856 | 0.98636
| [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification) | 2019 | build a model that recognizes toxicity and minimizes this type of unintended bias with respect to mentions of identities. You'll be using a dataset labeled for identity mentions and optimizing a metric designed to measure unintended bias. | Civil Comments | `unbiased` | 0.94734 | 0.93639
| [Jigsaw Multilingual Toxic Comment Classification](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification) | 2020 | build effective multilingual models | Wikipedia Comments + Civil Comments | `multilingual` | 0.9536 | 0.91655*
*Score not directly comparable since it is obtained on the validation set provided and not on the test set. To update when the test labels are made available.
It is also noteworthy to mention that the top leadearboard scores have been achieved using model ensembles. The purpose of this library was to build something user-friendly and straightforward to use.
## Limitations and ethical considerations
If words that are associated with swearing, insults or profanity are present in a comment, it is likely that it will be classified as toxic, regardless of the tone or the intent of the author e.g. humorous/self-deprecating. This could present some biases towards already vulnerable minority groups.
The intended use of this library is for research purposes, fine-tuning on carefully constructed datasets that reflect real world demographics and/or to aid content moderators in flagging out harmful content quicker.
Some useful resources about the risk of different biases in toxicity or hate speech detection are:
- [The Risk of Racial Bias in Hate Speech Detection](https://homes.cs.washington.edu/~msap/pdfs/sap2019risk.pdf)
- [Automated Hate Speech Detection and the Problem of Offensive Language](https://arxiv.org/pdf/1703.04009.pdf%201.pdf)
- [Racial Bias in Hate Speech and Abusive Language Detection Datasets](https://arxiv.org/pdf/1905.12516.pdf)
## Quick prediction
The `multilingual` model has been trained on 7 different languages so it should only be tested on: `english`, `french`, `spanish`, `italian`, `portuguese`, `turkish` or `russian`.
```bash
# install detoxify
pip install detoxify
```
```python
from detoxify import Detoxify
# each model takes in either a string or a list of strings
results = Detoxify('original').predict('example text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# optional to display results nicely (will need to pip install pandas)
import pandas as pd
print(pd.DataFrame(results, index=input_text).round(5))
```
For more details check the Prediction section.
## Labels
All challenges have a toxicity label. The toxicity labels represent the aggregate ratings of up to 10 annotators according the following schema:
- **Very Toxic** (a very hateful, aggressive, or disrespectful comment that is very likely to make you leave a discussion or give up on sharing your perspective)
- **Toxic** (a rude, disrespectful, or unreasonable comment that is somewhat likely to make you leave a discussion or give up on sharing your perspective)
- **Hard to Say**
- **Not Toxic**
More information about the labelling schema can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Toxic Comment Classification Challenge
This challenge includes the following labels:
- `toxic`
- `severe_toxic`
- `obscene`
- `threat`
- `insult`
- `identity_hate`
### Jigsaw Unintended Bias in Toxicity Classification
This challenge has 2 types of labels: the main toxicity labels and some additional identity labels that represent the identities mentioned in the comments.
Only identities with more than 500 examples in the test set (combined public and private) are included during training as additional labels and in the evaluation calculation.
- `toxicity`
- `severe_toxicity`
- `obscene`
- `threat`
- `insult`
- `identity_attack`
- `sexual_explicit`
Identity labels used:
- `male`
- `female`
- `homosexual_gay_or_lesbian`
- `christian`
- `jewish`
- `muslim`
- `black`
- `white`
- `psychiatric_or_mental_illness`
A complete list of all the identity labels available can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Jigsaw Multilingual Toxic Comment Classification
Since this challenge combines the data from the previous 2 challenges, it includes all labels from above, however the final evaluation is only on:
- `toxicity`
## How to run
First, install dependencies
```bash
# clone project
git clone https://github.com/unitaryai/detoxify
# create virtual env
python3 -m venv toxic-env
source toxic-env/bin/activate
# install project
pip install -e detoxify
cd detoxify
# for training
pip install -r requirements.txt
```
## Prediction
Trained models summary:
|Model name| Transformer type| Data from
|:--:|:--:|:--:|
|`original`| `bert-base-uncased` | Toxic Comment Classification Challenge
|`unbiased`| `roberta-base`| Unintended Bias in Toxicity Classification
|`multilingual`| `xlm-roberta-base`| Multilingual Toxic Comment Classification
For a quick prediction can run the example script on a comment directly or from a txt containing a list of comments.
```bash
# load model via torch.hub
python run_prediction.py --input 'example' --model_name original
# load model from from checkpoint path
python run_prediction.py --input 'example' --from_ckpt_path model_path
# save results to a .csv file
python run_prediction.py --input test_set.txt --model_name original --save_to results.csv
# to see usage
python run_prediction.py --help
```
Checkpoints can be downloaded from the latest release or via the Pytorch hub API with the following names:
- `toxic_bert`
- `unbiased_toxic_roberta`
- `multilingual_toxic_xlm_r`
```bash
model = torch.hub.load('unitaryai/detoxify','toxic_bert')
```
Importing detoxify in python:
```python
from detoxify import Detoxify
results = Detoxify('original').predict('some text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# to display results nicely
import pandas as pd
print(pd.DataFrame(results,index=input_text).round(5))
```
## Training
If you do not already have a Kaggle account:
- you need to create one to be able to download the data
- go to My Account and click on Create New API Token - this will download a kaggle.json file
- make sure this file is located in ~/.kaggle
```bash
# create data directory
mkdir jigsaw_data
cd jigsaw_data
# download data
kaggle competitions download -c jigsaw-toxic-comment-classification-challenge
kaggle competitions download -c jigsaw-unintended-bias-in-toxicity-classification
kaggle competitions download -c jigsaw-multilingual-toxic-comment-classification
```
## Start Training
### Toxic Comment Classification Challenge
```bash
python create_val_set.py
python train.py --config configs/Toxic_comment_classification_BERT.json
```
### Unintended Bias in Toxicicity Challenge
```bash
python train.py --config configs/Unintended_bias_toxic_comment_classification_RoBERTa.json
```
### Multilingual Toxic Comment Classification
This is trained in 2 stages. First, train on all available data, and second, train only on the translated versions of the first challenge.
The [translated data](https://www.kaggle.com/miklgr500/jigsaw-train-multilingual-coments-google-api) can be downloaded from Kaggle in french, spanish, italian, portuguese, turkish, and russian (the languages available in the test set).
```bash
# stage 1
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR.json
# stage 2
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR_stage2.json
```
### Monitor progress with tensorboard
```bash
tensorboard --logdir=./saved
```
## Model Evaluation
### Toxic Comment Classification Challenge
This challenge is evaluated on the mean AUC score of all the labels.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Unintended Bias in Toxicicity Challenge
This challenge is evaluated on a novel bias metric that combines different AUC scores to balance overall performance. More information on this metric [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/overview/evaluation).
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
# to get the final bias metric
python model_eval/compute_bias_metric.py
```
### Multilingual Toxic Comment Classification
This challenge is evaluated on the AUC score of the main toxic label.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Citation
```
@misc{Detoxify,
title={Detoxify},
author={Hanu, Laura and {Unitary team}},
howpublished={Github. https://github.com/unitaryai/detoxify},
year={2020}
}
```
|
unitary/toxic-bert | 2021-06-07T15:20:33.000Z | [
"pytorch",
"jax",
"bert",
"text-classification",
"arxiv:1703.04009",
"arxiv:1905.12516",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.txt"
] | unitary | 2,303 | transformers |
<div align="center">
**⚠️ Disclaimer:**
The huggingface models currently give different results to the detoxify library (see issue [here](https://github.com/unitaryai/detoxify/issues/15)). For the most up to date models we recommend using the models from https://github.com/unitaryai/detoxify
# 🙊 Detoxify
## Toxic Comment Classification with ⚡ Pytorch Lightning and 🤗 Transformers
</div>
## Description
Trained models & code to predict toxic comments on 3 Jigsaw challenges: Toxic comment classification, Unintended Bias in Toxic comments, Multilingual toxic comment classification.
Built by [Laura Hanu](https://laurahanu.github.io/) at [Unitary](https://www.unitary.ai/), where we are working to stop harmful content online by interpreting visual content in context.
Dependencies:
- For inference:
- 🤗 Transformers
- ⚡ Pytorch lightning
- For training will also need:
- Kaggle API (to download data)
| Challenge | Year | Goal | Original Data Source | Detoxify Model Name | Top Kaggle Leaderboard Score | Detoxify Score
|-|-|-|-|-|-|-|
| [Toxic Comment Classification Challenge](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge) | 2018 | build a multi-headed model that’s capable of detecting different types of of toxicity like threats, obscenity, insults, and identity-based hate. | Wikipedia Comments | `original` | 0.98856 | 0.98636
| [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification) | 2019 | build a model that recognizes toxicity and minimizes this type of unintended bias with respect to mentions of identities. You'll be using a dataset labeled for identity mentions and optimizing a metric designed to measure unintended bias. | Civil Comments | `unbiased` | 0.94734 | 0.93639
| [Jigsaw Multilingual Toxic Comment Classification](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification) | 2020 | build effective multilingual models | Wikipedia Comments + Civil Comments | `multilingual` | 0.9536 | 0.91655*
*Score not directly comparable since it is obtained on the validation set provided and not on the test set. To update when the test labels are made available.
It is also noteworthy to mention that the top leadearboard scores have been achieved using model ensembles. The purpose of this library was to build something user-friendly and straightforward to use.
## Limitations and ethical considerations
If words that are associated with swearing, insults or profanity are present in a comment, it is likely that it will be classified as toxic, regardless of the tone or the intent of the author e.g. humorous/self-deprecating. This could present some biases towards already vulnerable minority groups.
The intended use of this library is for research purposes, fine-tuning on carefully constructed datasets that reflect real world demographics and/or to aid content moderators in flagging out harmful content quicker.
Some useful resources about the risk of different biases in toxicity or hate speech detection are:
- [The Risk of Racial Bias in Hate Speech Detection](https://homes.cs.washington.edu/~msap/pdfs/sap2019risk.pdf)
- [Automated Hate Speech Detection and the Problem of Offensive Language](https://arxiv.org/pdf/1703.04009.pdf%201.pdf)
- [Racial Bias in Hate Speech and Abusive Language Detection Datasets](https://arxiv.org/pdf/1905.12516.pdf)
## Quick prediction
The `multilingual` model has been trained on 7 different languages so it should only be tested on: `english`, `french`, `spanish`, `italian`, `portuguese`, `turkish` or `russian`.
```bash
# install detoxify
pip install detoxify
```
```python
from detoxify import Detoxify
# each model takes in either a string or a list of strings
results = Detoxify('original').predict('example text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# optional to display results nicely (will need to pip install pandas)
import pandas as pd
print(pd.DataFrame(results, index=input_text).round(5))
```
For more details check the Prediction section.
## Labels
All challenges have a toxicity label. The toxicity labels represent the aggregate ratings of up to 10 annotators according the following schema:
- **Very Toxic** (a very hateful, aggressive, or disrespectful comment that is very likely to make you leave a discussion or give up on sharing your perspective)
- **Toxic** (a rude, disrespectful, or unreasonable comment that is somewhat likely to make you leave a discussion or give up on sharing your perspective)
- **Hard to Say**
- **Not Toxic**
More information about the labelling schema can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Toxic Comment Classification Challenge
This challenge includes the following labels:
- `toxic`
- `severe_toxic`
- `obscene`
- `threat`
- `insult`
- `identity_hate`
### Jigsaw Unintended Bias in Toxicity Classification
This challenge has 2 types of labels: the main toxicity labels and some additional identity labels that represent the identities mentioned in the comments.
Only identities with more than 500 examples in the test set (combined public and private) are included during training as additional labels and in the evaluation calculation.
- `toxicity`
- `severe_toxicity`
- `obscene`
- `threat`
- `insult`
- `identity_attack`
- `sexual_explicit`
Identity labels used:
- `male`
- `female`
- `homosexual_gay_or_lesbian`
- `christian`
- `jewish`
- `muslim`
- `black`
- `white`
- `psychiatric_or_mental_illness`
A complete list of all the identity labels available can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Jigsaw Multilingual Toxic Comment Classification
Since this challenge combines the data from the previous 2 challenges, it includes all labels from above, however the final evaluation is only on:
- `toxicity`
## How to run
First, install dependencies
```bash
# clone project
git clone https://github.com/unitaryai/detoxify
# create virtual env
python3 -m venv toxic-env
source toxic-env/bin/activate
# install project
pip install -e detoxify
cd detoxify
# for training
pip install -r requirements.txt
```
## Prediction
Trained models summary:
|Model name| Transformer type| Data from
|:--:|:--:|:--:|
|`original`| `bert-base-uncased` | Toxic Comment Classification Challenge
|`unbiased`| `roberta-base`| Unintended Bias in Toxicity Classification
|`multilingual`| `xlm-roberta-base`| Multilingual Toxic Comment Classification
For a quick prediction can run the example script on a comment directly or from a txt containing a list of comments.
```bash
# load model via torch.hub
python run_prediction.py --input 'example' --model_name original
# load model from from checkpoint path
python run_prediction.py --input 'example' --from_ckpt_path model_path
# save results to a .csv file
python run_prediction.py --input test_set.txt --model_name original --save_to results.csv
# to see usage
python run_prediction.py --help
```
Checkpoints can be downloaded from the latest release or via the Pytorch hub API with the following names:
- `toxic_bert`
- `unbiased_toxic_roberta`
- `multilingual_toxic_xlm_r`
```bash
model = torch.hub.load('unitaryai/detoxify','toxic_bert')
```
Importing detoxify in python:
```python
from detoxify import Detoxify
results = Detoxify('original').predict('some text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# to display results nicely
import pandas as pd
print(pd.DataFrame(results,index=input_text).round(5))
```
## Training
If you do not already have a Kaggle account:
- you need to create one to be able to download the data
- go to My Account and click on Create New API Token - this will download a kaggle.json file
- make sure this file is located in ~/.kaggle
```bash
# create data directory
mkdir jigsaw_data
cd jigsaw_data
# download data
kaggle competitions download -c jigsaw-toxic-comment-classification-challenge
kaggle competitions download -c jigsaw-unintended-bias-in-toxicity-classification
kaggle competitions download -c jigsaw-multilingual-toxic-comment-classification
```
## Start Training
### Toxic Comment Classification Challenge
```bash
python create_val_set.py
python train.py --config configs/Toxic_comment_classification_BERT.json
```
### Unintended Bias in Toxicicity Challenge
```bash
python train.py --config configs/Unintended_bias_toxic_comment_classification_RoBERTa.json
```
### Multilingual Toxic Comment Classification
This is trained in 2 stages. First, train on all available data, and second, train only on the translated versions of the first challenge.
The [translated data](https://www.kaggle.com/miklgr500/jigsaw-train-multilingual-coments-google-api) can be downloaded from Kaggle in french, spanish, italian, portuguese, turkish, and russian (the languages available in the test set).
```bash
# stage 1
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR.json
# stage 2
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR_stage2.json
```
### Monitor progress with tensorboard
```bash
tensorboard --logdir=./saved
```
## Model Evaluation
### Toxic Comment Classification Challenge
This challenge is evaluated on the mean AUC score of all the labels.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Unintended Bias in Toxicicity Challenge
This challenge is evaluated on a novel bias metric that combines different AUC scores to balance overall performance. More information on this metric [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/overview/evaluation).
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
# to get the final bias metric
python model_eval/compute_bias_metric.py
```
### Multilingual Toxic Comment Classification
This challenge is evaluated on the AUC score of the main toxic label.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Citation
```
@misc{Detoxify,
title={Detoxify},
author={Hanu, Laura and {Unitary team}},
howpublished={Github. https://github.com/unitaryai/detoxify},
year={2020}
}
```
|
unitary/unbiased-toxic-roberta | 2021-06-07T15:20:21.000Z | [
"pytorch",
"jax",
"roberta",
"text-classification",
"arxiv:1703.04009",
"arxiv:1905.12516",
"transformers"
] | text-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | unitary | 101 | transformers | <div align="center">
**⚠️ Disclaimer:**
The huggingface models currently give different results to the detoxify library (see issue [here](https://github.com/unitaryai/detoxify/issues/15)). For the most up to date models we recommend using the models from https://github.com/unitaryai/detoxify
# 🙊 Detoxify
## Toxic Comment Classification with ⚡ Pytorch Lightning and 🤗 Transformers
</div>
## Description
Trained models & code to predict toxic comments on 3 Jigsaw challenges: Toxic comment classification, Unintended Bias in Toxic comments, Multilingual toxic comment classification.
Built by [Laura Hanu](https://laurahanu.github.io/) at [Unitary](https://www.unitary.ai/), where we are working to stop harmful content online by interpreting visual content in context.
Dependencies:
- For inference:
- 🤗 Transformers
- ⚡ Pytorch lightning
- For training will also need:
- Kaggle API (to download data)
| Challenge | Year | Goal | Original Data Source | Detoxify Model Name | Top Kaggle Leaderboard Score | Detoxify Score
|-|-|-|-|-|-|-|
| [Toxic Comment Classification Challenge](https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge) | 2018 | build a multi-headed model that’s capable of detecting different types of of toxicity like threats, obscenity, insults, and identity-based hate. | Wikipedia Comments | `original` | 0.98856 | 0.98636
| [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification) | 2019 | build a model that recognizes toxicity and minimizes this type of unintended bias with respect to mentions of identities. You'll be using a dataset labeled for identity mentions and optimizing a metric designed to measure unintended bias. | Civil Comments | `unbiased` | 0.94734 | 0.93639
| [Jigsaw Multilingual Toxic Comment Classification](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification) | 2020 | build effective multilingual models | Wikipedia Comments + Civil Comments | `multilingual` | 0.9536 | 0.91655*
*Score not directly comparable since it is obtained on the validation set provided and not on the test set. To update when the test labels are made available.
It is also noteworthy to mention that the top leadearboard scores have been achieved using model ensembles. The purpose of this library was to build something user-friendly and straightforward to use.
## Limitations and ethical considerations
If words that are associated with swearing, insults or profanity are present in a comment, it is likely that it will be classified as toxic, regardless of the tone or the intent of the author e.g. humorous/self-deprecating. This could present some biases towards already vulnerable minority groups.
The intended use of this library is for research purposes, fine-tuning on carefully constructed datasets that reflect real world demographics and/or to aid content moderators in flagging out harmful content quicker.
Some useful resources about the risk of different biases in toxicity or hate speech detection are:
- [The Risk of Racial Bias in Hate Speech Detection](https://homes.cs.washington.edu/~msap/pdfs/sap2019risk.pdf)
- [Automated Hate Speech Detection and the Problem of Offensive Language](https://arxiv.org/pdf/1703.04009.pdf%201.pdf)
- [Racial Bias in Hate Speech and Abusive Language Detection Datasets](https://arxiv.org/pdf/1905.12516.pdf)
## Quick prediction
The `multilingual` model has been trained on 7 different languages so it should only be tested on: `english`, `french`, `spanish`, `italian`, `portuguese`, `turkish` or `russian`.
```bash
# install detoxify
pip install detoxify
```
```python
from detoxify import Detoxify
# each model takes in either a string or a list of strings
results = Detoxify('original').predict('example text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# optional to display results nicely (will need to pip install pandas)
import pandas as pd
print(pd.DataFrame(results, index=input_text).round(5))
```
For more details check the Prediction section.
## Labels
All challenges have a toxicity label. The toxicity labels represent the aggregate ratings of up to 10 annotators according the following schema:
- **Very Toxic** (a very hateful, aggressive, or disrespectful comment that is very likely to make you leave a discussion or give up on sharing your perspective)
- **Toxic** (a rude, disrespectful, or unreasonable comment that is somewhat likely to make you leave a discussion or give up on sharing your perspective)
- **Hard to Say**
- **Not Toxic**
More information about the labelling schema can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Toxic Comment Classification Challenge
This challenge includes the following labels:
- `toxic`
- `severe_toxic`
- `obscene`
- `threat`
- `insult`
- `identity_hate`
### Jigsaw Unintended Bias in Toxicity Classification
This challenge has 2 types of labels: the main toxicity labels and some additional identity labels that represent the identities mentioned in the comments.
Only identities with more than 500 examples in the test set (combined public and private) are included during training as additional labels and in the evaluation calculation.
- `toxicity`
- `severe_toxicity`
- `obscene`
- `threat`
- `insult`
- `identity_attack`
- `sexual_explicit`
Identity labels used:
- `male`
- `female`
- `homosexual_gay_or_lesbian`
- `christian`
- `jewish`
- `muslim`
- `black`
- `white`
- `psychiatric_or_mental_illness`
A complete list of all the identity labels available can be found [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data).
### Jigsaw Multilingual Toxic Comment Classification
Since this challenge combines the data from the previous 2 challenges, it includes all labels from above, however the final evaluation is only on:
- `toxicity`
## How to run
First, install dependencies
```bash
# clone project
git clone https://github.com/unitaryai/detoxify
# create virtual env
python3 -m venv toxic-env
source toxic-env/bin/activate
# install project
pip install -e detoxify
cd detoxify
# for training
pip install -r requirements.txt
```
## Prediction
Trained models summary:
|Model name| Transformer type| Data from
|:--:|:--:|:--:|
|`original`| `bert-base-uncased` | Toxic Comment Classification Challenge
|`unbiased`| `roberta-base`| Unintended Bias in Toxicity Classification
|`multilingual`| `xlm-roberta-base`| Multilingual Toxic Comment Classification
For a quick prediction can run the example script on a comment directly or from a txt containing a list of comments.
```bash
# load model via torch.hub
python run_prediction.py --input 'example' --model_name original
# load model from from checkpoint path
python run_prediction.py --input 'example' --from_ckpt_path model_path
# save results to a .csv file
python run_prediction.py --input test_set.txt --model_name original --save_to results.csv
# to see usage
python run_prediction.py --help
```
Checkpoints can be downloaded from the latest release or via the Pytorch hub API with the following names:
- `toxic_bert`
- `unbiased_toxic_roberta`
- `multilingual_toxic_xlm_r`
```bash
model = torch.hub.load('unitaryai/detoxify','toxic_bert')
```
Importing detoxify in python:
```python
from detoxify import Detoxify
results = Detoxify('original').predict('some text')
results = Detoxify('unbiased').predict(['example text 1','example text 2'])
results = Detoxify('multilingual').predict(['example text','exemple de texte','texto de ejemplo','testo di esempio','texto de exemplo','örnek metin','пример текста'])
# to display results nicely
import pandas as pd
print(pd.DataFrame(results,index=input_text).round(5))
```
## Training
If you do not already have a Kaggle account:
- you need to create one to be able to download the data
- go to My Account and click on Create New API Token - this will download a kaggle.json file
- make sure this file is located in ~/.kaggle
```bash
# create data directory
mkdir jigsaw_data
cd jigsaw_data
# download data
kaggle competitions download -c jigsaw-toxic-comment-classification-challenge
kaggle competitions download -c jigsaw-unintended-bias-in-toxicity-classification
kaggle competitions download -c jigsaw-multilingual-toxic-comment-classification
```
## Start Training
### Toxic Comment Classification Challenge
```bash
python create_val_set.py
python train.py --config configs/Toxic_comment_classification_BERT.json
```
### Unintended Bias in Toxicicity Challenge
```bash
python train.py --config configs/Unintended_bias_toxic_comment_classification_RoBERTa.json
```
### Multilingual Toxic Comment Classification
This is trained in 2 stages. First, train on all available data, and second, train only on the translated versions of the first challenge.
The [translated data](https://www.kaggle.com/miklgr500/jigsaw-train-multilingual-coments-google-api) can be downloaded from Kaggle in french, spanish, italian, portuguese, turkish, and russian (the languages available in the test set).
```bash
# stage 1
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR.json
# stage 2
python train.py --config configs/Multilingual_toxic_comment_classification_XLMR_stage2.json
```
### Monitor progress with tensorboard
```bash
tensorboard --logdir=./saved
```
## Model Evaluation
### Toxic Comment Classification Challenge
This challenge is evaluated on the mean AUC score of all the labels.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Unintended Bias in Toxicicity Challenge
This challenge is evaluated on a novel bias metric that combines different AUC scores to balance overall performance. More information on this metric [here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/overview/evaluation).
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
# to get the final bias metric
python model_eval/compute_bias_metric.py
```
### Multilingual Toxic Comment Classification
This challenge is evaluated on the AUC score of the main toxic label.
```bash
python evaluate.py --checkpoint saved/lightning_logs/checkpoints/example_checkpoint.pth --test_csv test.csv
```
### Citation
```
@misc{Detoxify,
title={Detoxify},
author={Hanu, Laura and {Unitary team}},
howpublished={Github. https://github.com/unitaryai/detoxify},
year={2020}
}
```
|
unknownTransformer/wav2vec2-large-xlsr-german | 2021-05-11T08:43:09.000Z | [
"pytorch",
"wav2vec2",
"transformers"
] | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | unknownTransformer | 46 | transformers | Bad Modell for Research Purposes! |
|
upboys/SemanticSearch | 2021-02-11T07:37:53.000Z | [] | [
".gitattributes"
] | upboys | 0 | |||
upendra1881/bert_large_uncase_tweet_qa | 2021-04-27T13:51:28.000Z | [] | [
".gitattributes",
"README.md"
] | upendra1881 | 0 | |||
upendra1881/bert_tweetQA | 2021-04-27T11:50:30.000Z | [] | [
".gitattributes"
] | upendra1881 | 0 | |||
urduhack/roberta-urdu-small | 2021-05-20T22:52:23.000Z | [
"pytorch",
"jax",
"roberta",
"masked-lm",
"ur",
"transformers",
"roberta-urdu-small",
"urdu",
"license:mit",
"fill-mask"
] | fill-mask | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"training_args.bin",
"vocab.json"
] | urduhack | 197 | transformers | ---
language: ur
thumbnail: https://raw.githubusercontent.com/urduhack/urduhack/master/docs/_static/urduhack.png
tags:
- roberta-urdu-small
- urdu
- transformers
license: mit
---
## roberta-urdu-small
[](https://github.com/urduhack/urduhack/blob/master/LICENSE)
### Overview
**Language model:** roberta-urdu-small
**Model size:** 125M
**Language:** Urdu
**Training data:** News data from urdu news resources in Pakistan
### About roberta-urdu-small
roberta-urdu-small is a language model for urdu language.
```
from transformers import pipeline
fill_mask = pipeline("fill-mask", model="urduhack/roberta-urdu-small", tokenizer="urduhack/roberta-urdu-small")
```
## Training procedure
roberta-urdu-small was trained on urdu news corpus. Training data was normalized using normalization module from
urduhack to eliminate characters from other languages like arabic.
### About Urduhack
Urduhack is a Natural Language Processing (NLP) library for urdu language.
Github: https://github.com/urduhack/urduhack
|
userSE075/news_class | 2021-04-21T11:58:25.000Z | [] | [
".gitattributes"
] | userSE075 | 0 | |||
ushikado/yuyuyui-chatbot | 2021-05-23T13:27:10.000Z | [
"pytorch",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"ja",
"transformers",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"added_tokens.json",
"config.json",
"flax_model.msgpack",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tokenizer_config.json"
] | ushikado | 42 | transformers | ---
language: ja
inference: false
---
# yuyuyui-chatbot
This model is based on [rinna/japanese-gpt2-medium](https://huggingface.co/rinna/japanese-gpt2-medium) and finetuned on Yuyuyui scenario corpus.
## Usage
The model takes a sequence of utterances (context) to generate a subsequent utterance (response). Each utterance begins with a **character token** and ends with an **EOS token**. Use the unspecified character token `<某>` for user inputs.
Put a character token after your question or query to generate a response from a specific character. In this case, make sure that an EOS token is not appended automatically by the tokenizer. Otherwise the model will interpret the trailing EOS as an empty utterance and try to add another random character token.
Simple example:
```python
from transformers import T5Tokenizer, AutoModelForCausalLM
tokenizer = T5Tokenizer.from_pretrained("ushikado/yuyuyui-chatbot")
model = AutoModelForCausalLM.from_pretrained("ushikado/yuyuyui-chatbot")
query_text = "<某>神樹様について教えてください。</s><上里 ひなた>"
input_tensor = tokenizer.encode(query_text, add_special_tokens=False, return_tensors="pt")
output_list = model.generate(input_tensor, max_length=100, do_sample=True, pad_token_id=tokenizer.eos_token_id)
output_text = tokenizer.decode(output_list[0])
print(output_text)
"""
<某> 神樹様について教えてください。</s> <上里 ひなた> 造反神は、神樹様の分裂を煽り出して、神樹様の中の一体感を高める存在です。</s>
"""
```
Accumulate dialog history to make responses more context-aware:
```python
class Interlocutor():
def __init__(self, tokenizer, model, character_token, max_context_length=512, max_response_length=128):
self.tokenizer = tokenizer
self.model = model
self.character_token = character_token
self.max_context_length = max_context_length
self.max_response_length = max_response_length
self.context = ""
return
def generate(self, query):
nanigashi = self.tokenizer.additional_special_tokens[0]
nanigashi_id = self.tokenizer.additional_special_tokens_ids[0]
self.context += nanigashi + query + self.tokenizer.eos_token + self.character_token
context_tensor = self.tokenizer.encode(self.context, add_special_tokens=False, return_tensors="pt")
context_length = context_tensor.size()[-1]
if self.max_context_length < context_length:
context_tensor = context_tensor.narrow(1, context_length - self.max_context_length, self.max_context_length)
context_length = context_tensor.size()[-1]
max_length = context_length + self.max_response_length
context_tensor = self.model.generate(context_tensor, do_sample=True, max_length=max_length,
pad_token_id=self.tokenizer.eos_token_id)
self.context = re.sub(self.tokenizer.eos_token, "", self.tokenizer.decode(context_tensor[0]))
response = self.context[self.context.rindex(self.character_token) + len(self.character_token) : ].strip()
print(response)
interlocutor = Interlocutor(tokenizer, model, "<加賀城 雀>")
interlocutor.generate("何しようかな。")
"""
そうだなぁ。せっかく徳島に来たんだから、何か食べたいよなー。</s>
"""
interlocutor.generate("例えば？")
"""
スパムとかいう高級料理はちょっとなぁ。あとは可愛い雑貨とか、おやつとか。</s>
"""
interlocutor.generate("徳島ラーメンじゃないの？")
"""
あー、確か徳島ラーメンってのがあって、それも美味しいんだよね。</s>
"""
interlocutor.generate("ここから近いお店があるんだって。行ってみよう！")
"""
わー! 何だか賑やかでいい感じだね。</s>
"""
interlocutor.generate("さっそく注文するね。")
"""
んー! ずっーと揚げ鶏が好きだったけど、今日は初めてまるまる鶏肉を注文してみるよ。</s>
"""
print(interlocutor.context)
"""
<某> 何しようかな。</s> <加賀城 雀> そうだなぁ。せっかく徳島に来たんだから、何か食べたいよなー。</s> <某> 例えば?</s> <加賀城 雀> スパムとかいう高級料理はちょっとなぁ。あとは可愛い雑貨とか、おやつとか。</s> <某> 徳島ラーメンじゃないの?</s> <加賀城 雀> あー、確か徳島ラーメンってのがあって、それも美味しいんだよね。</s> <某> ここから近いお店があるんだって。行ってみよう!</s> <加賀城 雀> わー! 何だか賑やかでいい感じだね。</s> <某> さっそく注文するね。</s> <加賀城 雀> んー! ずっーと揚げ鶏が好きだったけど、今日は初めてまるまる鶏肉を注文してみるよ。</s>
"""
```
## List of character tokens
`<某>` is _unspecified (nanigashi)_. Use for user inputs or mobs.
```plain
<某>
<結城 友奈>
<東郷 美森>
<犬吠埼 風>
<犬吠埼 樹>
<三好 夏凜>
<乃木 園子>
<鷲尾 須美>
<三ノ輪 銀>
<乃木 若葉>
<上里 ひなた>
<土居 球子>
<伊予島 杏>
<郡 千景>
<高嶋 友奈>
<白鳥 歌野>
<藤森 水都>
<秋原 雪花>
<古波蔵 棗>
<楠 芽吹>
<加賀城 雀>
<弥勒 夕海子>
<山伏 しずく>
<山伏 シズク>
<国土 亜耶>
<赤嶺 友奈>
<弥勒 蓮華>
<桐生 静>
<安芸 真鈴>
<花本 美佳>
```
## Licence
TBD. |
uw-hai/polyjuice | 2021-05-24T01:21:24.000Z | [
"pytorch",
"jax",
"gpt2",
"lm-head",
"causal-lm",
"en",
"transformers",
"counterfactual generation",
"text-generation"
] | text-generation | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"training_args.bin",
"vocab.json"
] | uw-hai | 317 | transformers | ---
language: "en"
tags:
- counterfactual generation
widget:
- text: "It is great for kids. <|perturb|> [negation] It [BLANK] great for kids. [SEP]"
---
# Polyjuice
## Model description
This is a ported version of [Polyjuice](https://homes.cs.washington.edu/~wtshuang/static/papers/2021-arxiv-polyjuice.pdf), the general-purpose counterfactual generator.
For more code release, please refer to [this github page](https://github.com/tongshuangwu/polyjuice).
#### How to use
```python
from transformers import AutoTokenizer, AutoModelWithLMHead
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
model_path = "uw-hai/polyjuice"
generator = pipeline("text-generation",
model=AutoModelForCausalLM.from_pretrained(model_path),
tokenizer=AutoTokenizer.from_pretrained(model_path),
framework="pt", device=0 if is_cuda else -1)
prompt_text = "A dog is embraced by the woman. <|perturb|> [negation] A dog is [BLANK] the woman."
generator(prompt_text, num_beams=3, num_return_sequences=3)
```
### BibTeX entry and citation info
```bibtex
@inproceedings{polyjuice:acl21,
title = "{P}olyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models",
author = "Tongshuang Wu and Marco Tulio Ribeiro and Jeffrey Heer and Daniel S. Weld",
booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics",
year = "2021",
publisher = "Association for Computational Linguistics"
``` |
vahmohh/t5-qag-base | 2021-03-01T12:42:47.000Z | [
"pytorch",
"t5",
"seq2seq",
"transformers",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"added_tokens.json",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tokenizer_config.json"
] | vahmohh | 7 | transformers | [www.github.com/vahmohh/masters-thesis](https://www.github.com/vahmohh/masters-thesis)
The model has been built upon the pre-trained T5 model by fine-tuning it on SQuAD dataset for the porpuse of automatic question and answer generation.
The following format should be used for generating questions.
```sh
generate question: domain_specific_text </sep> answer_1 </sep> answer_2 </sep> ... </sep> answer_n </end>
```
Output:
```sh
question_1 </sep> question_2 </sep> ... </sep> question_n </end>
```
The following format should be used for generating answers.
```sh
generate answer: domain_specific_text </end>
```
Output:
```sh
answer_1 </sep> answer_2 </sep> ... </sep> answer_n </end>
``` |
vaishnavi/indic-bert-512 | 2021-04-08T06:38:32.000Z | [
"pytorch",
"albert",
"en",
"dataset:AI4Bharat IndicNLP Corpora",
"transformers",
"license:mit"
] | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"spiece.model",
"spiece.vocab",
"tf_model.ckpt.data-00000-of-00001",
"tf_model.ckpt.index",
"tf_model.ckpt.meta"
] | vaishnavi | 6 | transformers | ---
language: en
license: mit
datasets:
- AI4Bharat IndicNLP Corpora
---
# IndicBERT
IndicBERT is a multilingual ALBERT model pretrained exclusively on 12 major Indian languages. It is pre-trained on our novel monolingual corpus of around 9 billion tokens and subsequently evaluated on a set of diverse tasks. IndicBERT has much fewer parameters than other multilingual models (mBERT, XLM-R etc.) while it also achieves a performance on-par or better than these models.
The 12 languages covered by IndicBERT are: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, Telugu.
The code can be found [here](https://github.com/divkakwani/indic-bert). For more information, checkout our [project page](https://indicnlp.ai4bharat.org/) or our [paper](https://indicnlp.ai4bharat.org/papers/arxiv2020_indicnlp_corpus.pdf).
## Pretraining Corpus
We pre-trained indic-bert on AI4Bharat's monolingual corpus. The corpus has the following distribution of languages:
| Language | as | bn | en | gu | hi | kn | |
| ----------------- | ------ | ------ | ------ | ------ | ------ | ------ | ------- |
| **No. of Tokens** | 36.9M | 815M | 1.34B | 724M | 1.84B | 712M | |
| **Language** | **ml** | **mr** | **or** | **pa** | **ta** | **te** | **all** |
| **No. of Tokens** | 767M | 560M | 104M | 814M | 549M | 671M | 8.9B |
## Evaluation Results
IndicBERT is evaluated on IndicGLUE and some additional tasks. The results are summarized below. For more details about the tasks, refer our [official repo](https://github.com/divkakwani/indic-bert)
#### IndicGLUE
Task | mBERT | XLM-R | IndicBERT
-----| ----- | ----- | ------
News Article Headline Prediction | 89.58 | 95.52 | **95.87**
Wikipedia Section Title Prediction| **73.66** | 66.33 | 73.31
Cloze-style multiple-choice QA | 39.16 | 27.98 | **41.87**
Article Genre Classification | 90.63 | 97.03 | **97.34**
Named Entity Recognition (F1-score) | **73.24** | 65.93 | 64.47
Cross-Lingual Sentence Retrieval Task | 21.46 | 13.74 | **27.12**
Average | 64.62 | 61.09 | **66.66**
#### Additional Tasks
Task | Task Type | mBERT | XLM-R | IndicBERT
-----| ----- | ----- | ------ | -----
BBC News Classification | Genre Classification | 60.55 | **75.52** | 74.60
IIT Product Reviews | Sentiment Analysis | 74.57 | **78.97** | 71.32
IITP Movie Reviews | Sentiment Analaysis | 56.77 | **61.61** | 59.03
Soham News Article | Genre Classification | 80.23 | **87.6** | 78.45
Midas Discourse | Discourse Analysis | 71.20 | **79.94** | 78.44
iNLTK Headlines Classification | Genre Classification | 87.95 | 93.38 | **94.52**
ACTSA Sentiment Analysis | Sentiment Analysis | 48.53 | 59.33 | **61.18**
Winograd NLI | Natural Language Inference | 56.34 | 55.87 | **56.34**
Choice of Plausible Alternative (COPA) | Natural Language Inference | 54.92 | 51.13 | **58.33**
Amrita Exact Paraphrase | Paraphrase Detection | **93.81** | 93.02 | 93.75
Amrita Rough Paraphrase | Paraphrase Detection | 83.38 | 82.20 | **84.33**
Average | | 69.84 | **74.42** | 73.66
\* Note: all models have been restricted to a max_seq_length of 128.
## Downloads
The model can be downloaded [here](https://storage.googleapis.com/ai4bharat-public-indic-nlp-corpora/models/indic-bert-v1.tar.gz). Both tf checkpoints and pytorch binaries are included in the archive. Alternatively, you can also download it from [Huggingface](https://huggingface.co/ai4bharat/indic-bert).
## Citing
If you are using any of the resources, please cite the following article:
```
@inproceedings{kakwani2020indicnlpsuite,
title={{IndicNLPSuite: Monolingual Corpora, Evaluation Benchmarks and Pre-trained Multilingual Language Models for Indian Languages}},
author={Divyanshu Kakwani and Anoop Kunchukuttan and Satish Golla and Gokul N.C. and Avik Bhattacharyya and Mitesh M. Khapra and Pratyush Kumar},
year={2020},
booktitle={Findings of EMNLP},
}
```
We would like to hear from you if:
- You are using our resources. Please let us know how you are putting these resources to use.
- You have any feedback on these resources.
## License
The IndicBERT code (and models) are released under the MIT License.
## Contributors
- Divyanshu Kakwani
- Anoop Kunchukuttan
- Gokul NC
- Satish Golla
- Avik Bhattacharyya
- Mitesh Khapra
- Pratyush Kumar
This work is the outcome of a volunteer effort as part of [AI4Bharat initiative](https://ai4bharat.org).
## Contact
- Anoop Kunchukuttan ([[email protected]](mailto:[email protected]))
- Mitesh Khapra ([[email protected]](mailto:[email protected]))
- Pratyush Kumar ([[email protected]](mailto:[email protected]))
|
|
valhalla/bart-large-finetuned-squadv1 | 2021-06-14T10:20:35.000Z | [
"pytorch",
"jax",
"bart",
"question-answering",
"dataset:squad",
"arxiv:1910.13461",
"transformers"
] | question-answering | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | valhalla | 100,670 | transformers | ---
datasets:
- squad
---
# BART-LARGE finetuned on SQuADv1
This is bart-large model finetuned on SQuADv1 dataset for question answering task
## Model details
BART was propsed in the [paper](https://arxiv.org/abs/1910.13461) **BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension**.
BART is a seq2seq model intended for both NLG and NLU tasks.
To use BART for question answering tasks, we feed the complete document into the encoder and decoder, and use the top
hidden state of the decoder as a representation for each
word. This representation is used to classify the token. As given in the paper bart-large achives comparable to ROBERTa on SQuAD.
Another notable thing about BART is that it can handle sequences with upto 1024 tokens.
| Param | #Value |
|---------------------|--------|
| encoder layers | 12 |
| decoder layers | 12 |
| hidden size | 4096 |
| num attetion heads | 16 |
| on disk size | 1.63GB |
## Model training
This model was trained on google colab v100 GPU.
You can find the fine-tuning colab here
[](https://colab.research.google.com/drive/1I5cK1M_0dLaf5xoewh6swcm5nAInfwHy?usp=sharing).
## Results
The results are actually slightly worse than given in the paper.
In the paper the authors mentioned that bart-large achieves 88.8 EM and 94.6 F1
| Metric | #Value |
|--------|--------|
| EM | 86.8022|
| F1 | 92.7342|
## Model in Action 🚀
```python3
from transformers import BartTokenizer, BartForQuestionAnswering
import torch
tokenizer = BartTokenizer.from_pretrained('valhalla/bart-large-finetuned-squadv1')
model = BartForQuestionAnswering.from_pretrained('valhalla/bart-large-finetuned-squadv1')
question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
encoding = tokenizer(question, text, return_tensors='pt')
input_ids = encoding['input_ids']
attention_mask = encoding['attention_mask']
start_scores, end_scores = model(input_ids, attention_mask=attention_mask, output_attentions=False)[:2]
all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1])
answer = tokenizer.convert_tokens_to_ids(answer.split())
answer = tokenizer.decode(answer)
#answer => 'a nice puppet'
```
> Created with ❤️ by Suraj Patil [](https://github.com/patil-suraj/)
[](https://twitter.com/psuraj28)
|
valhalla/distilbart-mnli-12-1 | 2021-06-14T10:27:55.000Z | [
"pytorch",
"jax",
"bart",
"text-classification",
"dataset:mnli",
"transformers",
"distilbart",
"distilbart-mnli",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | valhalla | 7,981 | transformers | ---
datasets:
- mnli
tags:
- distilbart
- distilbart-mnli
pipeline_tag: zero-shot-classification
---
# DistilBart-MNLI
distilbart-mnli is the distilled version of bart-large-mnli created using the **No Teacher Distillation** technique proposed for BART summarisation by Huggingface, [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#distilbart).
We just copy alternating layers from `bart-large-mnli` and finetune more on the same data.
| | matched acc | mismatched acc |
| ------------------------------------------------------------------------------------ | ----------- | -------------- |
| [bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) (baseline, 12-12) | 89.9 | 90.01 |
| [distilbart-mnli-12-1](https://huggingface.co/valhalla/distilbart-mnli-12-1) | 87.08 | 87.5 |
| [distilbart-mnli-12-3](https://huggingface.co/valhalla/distilbart-mnli-12-3) | 88.1 | 88.19 |
| [distilbart-mnli-12-6](https://huggingface.co/valhalla/distilbart-mnli-12-6) | 89.19 | 89.01 |
| [distilbart-mnli-12-9](https://huggingface.co/valhalla/distilbart-mnli-12-9) | 89.56 | 89.52 |
This is a very simple and effective technique, as we can see the performance drop is very little.
Detailed performace trade-offs will be posted in this [sheet](https://docs.google.com/spreadsheets/d/1dQeUvAKpScLuhDV1afaPJRRAE55s2LpIzDVA5xfqxvk/edit?usp=sharing).
## Fine-tuning
If you want to train these models yourself, clone the [distillbart-mnli repo](https://github.com/patil-suraj/distillbart-mnli) and follow the steps below
Clone and install transformers from source
```bash
git clone https://github.com/huggingface/transformers.git
pip install -qqq -U ./transformers
```
Download MNLI data
```bash
python transformers/utils/download_glue_data.py --data_dir glue_data --tasks MNLI
```
Create student model
```bash
python create_student.py \
--teacher_model_name_or_path facebook/bart-large-mnli \
--student_encoder_layers 12 \
--student_decoder_layers 6 \
--save_path student-bart-mnli-12-6 \
```
Start fine-tuning
```bash
python run_glue.py args.json
```
You can find the logs of these trained models in this [wandb project](https://wandb.ai/psuraj/distilbart-mnli). |
valhalla/distilbart-mnli-12-3 | 2021-06-14T10:29:48.000Z | [
"pytorch",
"jax",
"bart",
"text-classification",
"dataset:mnli",
"transformers",
"distilbart",
"distilbart-mnli",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | valhalla | 7,419 | transformers | ---
datasets:
- mnli
tags:
- distilbart
- distilbart-mnli
pipeline_tag: zero-shot-classification
---
# DistilBart-MNLI
distilbart-mnli is the distilled version of bart-large-mnli created using the **No Teacher Distillation** technique proposed for BART summarisation by Huggingface, [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#distilbart).
We just copy alternating layers from `bart-large-mnli` and finetune more on the same data.
| | matched acc | mismatched acc |
| ------------------------------------------------------------------------------------ | ----------- | -------------- |
| [bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) (baseline, 12-12) | 89.9 | 90.01 |
| [distilbart-mnli-12-1](https://huggingface.co/valhalla/distilbart-mnli-12-1) | 87.08 | 87.5 |
| [distilbart-mnli-12-3](https://huggingface.co/valhalla/distilbart-mnli-12-3) | 88.1 | 88.19 |
| [distilbart-mnli-12-6](https://huggingface.co/valhalla/distilbart-mnli-12-6) | 89.19 | 89.01 |
| [distilbart-mnli-12-9](https://huggingface.co/valhalla/distilbart-mnli-12-9) | 89.56 | 89.52 |
This is a very simple and effective technique, as we can see the performance drop is very little.
Detailed performace trade-offs will be posted in this [sheet](https://docs.google.com/spreadsheets/d/1dQeUvAKpScLuhDV1afaPJRRAE55s2LpIzDVA5xfqxvk/edit?usp=sharing).
## Fine-tuning
If you want to train these models yourself, clone the [distillbart-mnli repo](https://github.com/patil-suraj/distillbart-mnli) and follow the steps below
Clone and install transformers from source
```bash
git clone https://github.com/huggingface/transformers.git
pip install -qqq -U ./transformers
```
Download MNLI data
```bash
python transformers/utils/download_glue_data.py --data_dir glue_data --tasks MNLI
```
Create student model
```bash
python create_student.py \
--teacher_model_name_or_path facebook/bart-large-mnli \
--student_encoder_layers 12 \
--student_decoder_layers 6 \
--save_path student-bart-mnli-12-6 \
```
Start fine-tuning
```bash
python run_glue.py args.json
```
You can find the logs of these trained models in this [wandb project](https://wandb.ai/psuraj/distilbart-mnli). |
valhalla/distilbart-mnli-12-6 | 2021-06-14T10:32:03.000Z | [
"pytorch",
"jax",
"bart",
"text-classification",
"dataset:mnli",
"transformers",
"distilbart",
"distilbart-mnli",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | valhalla | 468 | transformers | ---
datasets:
- mnli
tags:
- distilbart
- distilbart-mnli
pipeline_tag: zero-shot-classification
---
# DistilBart-MNLI
distilbart-mnli is the distilled version of bart-large-mnli created using the **No Teacher Distillation** technique proposed for BART summarisation by Huggingface, [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#distilbart).
We just copy alternating layers from `bart-large-mnli` and finetune more on the same data.
| | matched acc | mismatched acc |
| ------------------------------------------------------------------------------------ | ----------- | -------------- |
| [bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) (baseline, 12-12) | 89.9 | 90.01 |
| [distilbart-mnli-12-1](https://huggingface.co/valhalla/distilbart-mnli-12-1) | 87.08 | 87.5 |
| [distilbart-mnli-12-3](https://huggingface.co/valhalla/distilbart-mnli-12-3) | 88.1 | 88.19 |
| [distilbart-mnli-12-6](https://huggingface.co/valhalla/distilbart-mnli-12-6) | 89.19 | 89.01 |
| [distilbart-mnli-12-9](https://huggingface.co/valhalla/distilbart-mnli-12-9) | 89.56 | 89.52 |
This is a very simple and effective technique, as we can see the performance drop is very little.
Detailed performace trade-offs will be posted in this [sheet](https://docs.google.com/spreadsheets/d/1dQeUvAKpScLuhDV1afaPJRRAE55s2LpIzDVA5xfqxvk/edit?usp=sharing).
## Fine-tuning
If you want to train these models yourself, clone the [distillbart-mnli repo](https://github.com/patil-suraj/distillbart-mnli) and follow the steps below
Clone and install transformers from source
```bash
git clone https://github.com/huggingface/transformers.git
pip install -qqq -U ./transformers
```
Download MNLI data
```bash
python transformers/utils/download_glue_data.py --data_dir glue_data --tasks MNLI
```
Create student model
```bash
python create_student.py \
--teacher_model_name_or_path facebook/bart-large-mnli \
--student_encoder_layers 12 \
--student_decoder_layers 6 \
--save_path student-bart-mnli-12-6 \
```
Start fine-tuning
```bash
python run_glue.py args.json
```
You can find the logs of these trained models in this [wandb project](https://wandb.ai/psuraj/distilbart-mnli). |
valhalla/distilbart-mnli-12-9 | 2021-06-14T10:34:58.000Z | [
"pytorch",
"jax",
"bart",
"text-classification",
"dataset:mnli",
"transformers",
"distilbart",
"distilbart-mnli",
"zero-shot-classification",
"pipeline_tag:zero-shot-classification"
] | zero-shot-classification | [
".gitattributes",
"README.md",
"config.json",
"flax_model.msgpack",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
] | valhalla | 6,366 | transformers | ---
datasets:
- mnli
tags:
- distilbart
- distilbart-mnli
pipeline_tag: zero-shot-classification
---
# DistilBart-MNLI
distilbart-mnli is the distilled version of bart-large-mnli created using the **No Teacher Distillation** technique proposed for BART summarisation by Huggingface, [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#distilbart).
We just copy alternating layers from `bart-large-mnli` and finetune more on the same data.
| | matched acc | mismatched acc |
| ------------------------------------------------------------------------------------ | ----------- | -------------- |
| [bart-large-mnli](https://huggingface.co/facebook/bart-large-mnli) (baseline, 12-12) | 89.9 | 90.01 |
| [distilbart-mnli-12-1](https://huggingface.co/valhalla/distilbart-mnli-12-1) | 87.08 | 87.5 |
| [distilbart-mnli-12-3](https://huggingface.co/valhalla/distilbart-mnli-12-3) | 88.1 | 88.19 |
| [distilbart-mnli-12-6](https://huggingface.co/valhalla/distilbart-mnli-12-6) | 89.19 | 89.01 |
| [distilbart-mnli-12-9](https://huggingface.co/valhalla/distilbart-mnli-12-9) | 89.56 | 89.52 |
This is a very simple and effective technique, as we can see the performance drop is very little.
Detailed performace trade-offs will be posted in this [sheet](https://docs.google.com/spreadsheets/d/1dQeUvAKpScLuhDV1afaPJRRAE55s2LpIzDVA5xfqxvk/edit?usp=sharing).
## Fine-tuning
If you want to train these models yourself, clone the [distillbart-mnli repo](https://github.com/patil-suraj/distillbart-mnli) and follow the steps below
Clone and install transformers from source
```bash
git clone https://github.com/huggingface/transformers.git
pip install -qqq -U ./transformers
```
Download MNLI data
```bash
python transformers/utils/download_glue_data.py --data_dir glue_data --tasks MNLI
```
Create student model
```bash
python create_student.py \
--teacher_model_name_or_path facebook/bart-large-mnli \
--student_encoder_layers 12 \
--student_decoder_layers 6 \
--save_path student-bart-mnli-12-6 \
```
Start fine-tuning
```bash
python run_glue.py args.json
```
You can find the logs of these trained models in this [wandb project](https://wandb.ai/psuraj/distilbart-mnli). |
valhalla/distilt5-qa-qg-hl-12-6 | 2020-10-26T18:28:26.000Z | [
"pytorch",
"t5",
"seq2seq",
"dataset:squad",
"transformers",
"question-generation",
"distilt5",
"distilt5-qg",
"license:mit",
"text2text-generation"
] | text2text-generation | [
".gitattributes",
"README.md",
"added_tokens.json",
"config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"spiece.model",
"tokenizer_config.json"
] | valhalla | 805 | transformers | ---
datasets:
- squad
tags:
- question-generation
- distilt5
- distilt5-qg
widget:
- text: "generate question: <hl> 42 <hl> is the answer to life, the universe and everything. </s>"
- text: "question: What is 42 context: 42 is the answer to life, the universe and everything. </s>"
license: "MIT"
---
## DistilT5 for question-generation
This is distilled version of [t5-base-qa-qg-hl](https://huggingface.co/valhalla/t5-base-qa-qg-hl) model trained for question answering and answer aware question generation tasks.
The model is distilled using the **No Teacher Distillation** method proposed by Huggingface, [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq#distilbart).
We just copy alternating layers from `t5-base-qa-qg-hl` and finetune more on the same data. Following table lists other distilled models and their metrics.
| Name | BLEU-4 | METEOR | ROUGE-L | QA-EM | QA-F1 |
|---------------------------------------------------------------------------------|---------|---------|---------|--------|--------|
| [distilt5-qg-hl-6-4](https://huggingface.co/valhalla/distilt5-qg-hl-6-4) | 18.4141 | 24.8417 | 40.3435 | - | - |
| [distilt5-qa-qg-hl-6-4](https://huggingface.co/valhalla/distilt5-qa-qg-hl-6-4) | 18.6493 | 24.9685 | 40.5605 | 76.13 | 84.659 |
| [distilt5-qg-hl-12-6](https://huggingface.co/valhalla/distilt5-qg-hl-12-6) | 20.5275 | 26.5010 | 43.2676 | - | - |
| [distilt5-qa-qg-hl-12-6](https://huggingface.co/valhalla/distilt5-qa-qg-hl-12-6)| 20.6109 | 26.4533 | 43.0895 | 81.61 | 89.831 |
You can play with the model using the inference API. Here's how you can use it
For QG
`generate question: <hl> 42 <hl> is the answer to life, the universe and everything.`
For QA
`question: What is 42 context: 42 is the answer to life, the universe and everything.`
For more deatils see [this](https://github.com/patil-suraj/question_generation) repo.
### Model in action 🚀
You'll need to clone the [repo](https://github.com/patil-suraj/question_generation).
[](https://colab.research.google.com/github/patil-suraj/question_generation/blob/master/question_generation.ipynb)
```python3
from pipelines import pipeline
nlp = pipeline("multitask-qa-qg", model="valhalla/distilt5-qa-qg-hl-12-6")
# to generate questions simply pass the text
nlp("42 is the answer to life, the universe and everything.")
=> [{'answer': '42', 'question': 'What is the answer to life, the universe and everything?'}]
# for qa pass a dict with "question" and "context"
nlp({
"question": "What is 42 ?",
"context": "42 is the answer to life, the universe and everything."
})
=> 'the answer to life, the universe and everything'
``` |
Subsets and Splits