--- language: zh 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 set of GPT2 models, except for GPT2-xlarge model, are pre-trained by [UER-py](https://github.com/dbiir/UER-py/), which is introduced in [this paper](https://arxiv.org/abs/1909.05658). The GPT2-xlarge model is pre-trained by [TencentPretrain](https://github.com/Tencent/TencentPretrain) introduced in [this paper](https://arxiv.org/abs/2212.06385), which inherits UER-py to support models with parameters above one billion, and extends it to a multimodal pre-training framework. Besides, the other models could also be pre-trained by TencentPretrain. The model is used to generate Chinese texts. You can download the set of Chinese GPT2 models either from the [UER-py Modelzoo page](https://github.com/dbiir/UER-py/wiki/Modelzoo), or via HuggingFace from the links below: | | Link | | ----------------- | :----------------------------: | | **GPT2-distil** | [**L=6/H=768**][distil] | | **GPT2** | [**L=12/H=768**][base] | | **GPT2-medium** | [**L=24/H=1024**][medium] | | **GPT2-large** | [**L=36/H=1280**][large] | | **GPT2-xlarge** | [**L=48/H=1600**][xlarge] | Note that the 6-layer model is called GPT2-distil model because it follows the configuration of [distilgpt2](https://huggingface.co/distilgpt2), and 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 (take the case of GPT2-distil): ```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 GPT2-xlarge model is pre-trained by [TencentPretrain](https://github.com/Tencent/TencentPretrain), and the others 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 1024. For the models pre-trained by UER-py, take the case of GPT2-distil 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 --data_processor 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 ``` 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 --data_processor lm ``` ``` python3 pretrain.py --dataset_path cluecorpussmall_lm_seq1024_dataset.pt \ --vocab_path models/google_zh_vocab.txt \ --pretrained_model_path models/cluecorpussmall_gpt2_distil_seq128_model.bin-1000000 \ --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 ``` 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 ``` For GPT2-xlarge model, we use TencetPretrain. 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 --data_processor lm ``` ``` deepspeed pretrain.py --deepspeed --deepspeed_config models/deepspeed_config.json \ --dataset_path corpora/cluecorpussmall_lm_seq128_dataset.pt \ --vocab_path models/google_zh_vocab.txt \ --config_path models/gpt2/xlarge_config.json \ --output_model_path models/cluecorpussmall_gpt2_xlarge_seq128 \ --world_size 8 --batch_size 64 \ --total_steps 1000000 --save_checkpoint_steps 100000 --report_steps 50000 \ --deepspeed_checkpoint_activations --deepspeed_checkpoint_layers_num 24 ``` Before stage2, we extract fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints: ``` python3 models/cluecorpussmall_gpt2_xlarge_seq128/zero_to_fp32.py models/cluecorpussmall_gpt2_xlarge_seq128/ \ models/cluecorpussmall_gpt2_xlarge_seq128.bin ``` 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 --data_processor lm ``` ``` deepspeed pretrain.py --deepspeed --deepspeed_config models/deepspeed_config.json \ --dataset_path corpora/cluecorpussmall_lm_seq1024_dataset.pt \ --vocab_path models/google_zh_vocab.txt \ --config_path models/gpt2/xlarge_config.json \ --pretrained_model_path models/cluecorpussmall_gpt2_xlarge_seq128.bin \ --output_model_path models/cluecorpussmall_gpt2_xlarge_seq1024_stage2 \ --world_size 8 --batch_size 16 --learning_rate 5e-5 \ --total_steps 250000 --save_checkpoint_steps 50000 --report_steps 10000 \ --deepspeed_checkpoint_activations --deepspeed_checkpoint_layers_num 6 ``` Then, we extract fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints: ``` python3 models/cluecorpussmall_gpt2_xlarge_seq1024_stage2/zero_to_fp32.py models/cluecorpussmall_gpt2_xlarge_seq1024_stage2/ \ models/cluecorpussmall_gpt2_xlarge_seq1024_stage2.bin ``` Finally, we convert the pre-trained model into Huggingface's format: ``` python3 scripts/convert_gpt2_from_tencentpretrain_to_huggingface.py --input_model_path models/cluecorpussmall_gpt2_xlarge_seq1024_stage2.bin \ --output_model_path pytorch_model.bin \ --layers_num 48 ``` ### 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} } @article{zhao2023tencentpretrain, title={TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities}, author={Zhao, Zhe and Li, Yudong and Hou, Cheng and Zhao, Jing and others}, journal={ACL 2023}, pages={217}, year={2023} ``` [distil]:https://huggingface.co/uer/gpt2-distil-chinese-cluecorpussmall [base]:https://huggingface.co/uer/gpt2-chinese-cluecorpussmall [medium]:https://huggingface.co/uer/gpt2-medium-chinese-cluecorpussmall [large]:https://huggingface.co/uer/gpt2-large-chinese-cluecorpussmall [xlarge]:https://huggingface.co/uer/gpt2-xlarge-chinese-cluecorpussmall