• This project provides a large-scale Cleaned Chinese conversation dataset and Chinese pre-training dialog models trained on this dataset, and more details refer to our arxiv paper. The code is adapted from TransferTransfo using Transformers of HuggingFace, which can be used for pre-training and fine-tuning.
• 本项目提供了一个大规模的经过系统清洗的中文对话数据集，并提供在此数据集上的对话预训练模型，更多信息可参考我们的论文。 本项目代码修改自TransferTransfo，使用HuggingFace Pytorch版的Transformers, 可用于预训练与微调。

• Dataset
• Pre-training
• Evaluation

In this project, we present a Large-scale Cleaned Chinese Conversation corpus (LCCC) consists of LCCC-base and LCCC-large. The LCCC-base is cleaner but smaller than LCCC-large. The quality of our dataset is ensured by a rigorous data cleaning pipeline, which is built based on a set of rules and learned filter trained on manually annotated dialogue pairs. The noises we consider include dirty words, sensitive words, special symbols, emoji, grammatical errors, and context-free conversations. The statistic is described below, and the vocabulary words are counted based on Jieba segmentation.

The LCCC-base is constructed based on Weibo Corpus, which we crawled from Weibo. The LCCC-large is built on several conversation datasets in addition to Weibo Corpus:

We present a series of generative pre-training models for Chinese dialogue which are first pre-trained on the Chinese novel dataset and then post-trained on our LCCC dataset. The architecture is modified based on TransferTransfo, but we removed the classification module.

Similar to TransferTransfo, we concatenate all history utterances into one sequence as the context, and the goal of our model is to generate a response to each context. As shown below, the input of our model consists of word embedding, speaker embedding, and position embedding of each word.

Install from the source codes:

git clone https://github.com/lemon234071/GPT-Chinese.git
cd GPT-Chinese
pip install -r requirements.txt 

Step 1: Prepare the data and the pre-trianed model (train data or fine-tune data, E.g., STC dataset)

wget https://coai-dataset.oss-cn-beijing.aliyuncs.com/STC-corpus.zip # Download the STC dataset and unzip into "data_path" dir (fine-tuning on STC)
wget https://coai-dataset.oss-cn-beijing.aliyuncs.com/GPT_LCCC-large.zip # Download the GPT<sub>LCCC-large</sub> weights file and unzip into "model_checkpoint" dir

Note: If the computer's memory is insufficient, you can process the file into txt format, use the "train_path" to load data in a distributed manner (the function we adapted from Internet), and you need to leave "data_path" empty. You can also use the "toy_data.json" in our repository for test.

Step 2: Train the model

python train.py --pretrained --model_checkpoint ./models/ --data_path data/STC.json  # Single GPU training
python -m torch.distributed.launch --nproc_per_node=8 train.py --pretrained --model_checkpoint ./models/ --data_path data/STC.json  # Training on 8 GPUs

Step 3: Inference mode

python infer.py --model_checkpoint ./models/ --datapath data/STC_test.json --out_path STC_result.txt  # Do Inference on a corpus
python interact.py --model_checkpoint ./models/  # Interact on the terminal

Training Arguments

Evaluation is performed on results generated by models fine-tuned on STC dataset. All samples are generated by Nucleus Sampling with threshold 0.9 and temperature 0.7.

We also employed 3 workers to evaluate 200 samples of each model about fluency, relevance, and diversity based on a 2/1/0 score schema. Refer to our paper for more details.

We randomly select 1000 generated samples, and make a statistic of toxic generation:

Single-turn generated examples from STC dataset (fine-tuned on STC):

Multi-turn generated examples from an interactive environment:

Yida Wang, Pei Ke, Yinhe Zheng, Kaili Huang, Xiaoyan Zhu, Minlie Huang

The LCCC dataset and the pre-trained models aim to facilitate research for conversation generation. The LCCC dataset provided in this repository is crawled from many sources including Weibo. This project has carried out a rigorous cleaning on it, but it is not guaranteed to completely clean out the inappropriate content, which does not represent the author's opinion. This repository contains only part of the modeling machinery needed actually to produce a model weight file in a running dialog. The decoding script provided in this project is only for the researcher to test the generation effect of the pre-trained model. We are not responsible for any generation from the 3rd party utilization of the pre-trained system.

We'll release our paper on arxiv soon. Please kindly cite our paper if you use the dataset or models in your research.