gpt2-chinese-cluecorpussmall

Note that the 6-layer model is called GPT2-distil model because it follows the configuration of 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):

>>> 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 is used as training data.

Training procedure

The GPT2-xlarge model is pre-trained by TencentPretrain, and the others are pre-trained by UER-py on Tencent Cloud. 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 models/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_model \
                      --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_model/zero_to_fp32.py models/cluecorpussmall_gp