README.md

Data Processing

1. Process Scientific Pre-training Dataset

1.1 Process Chinese Dataset

cd pretrain/chinese_process/
python collect.py

1.2 Process English Dataset

cd pretrain/english_process/
python collect.py

Note that the redpajama_train.json is obtained by running the following command:

cd pretrain
python download_from_hf.py

1.3 Combine All the Dataset

First of all, get into the folder train, and run combine the chinese and english datasets and get the train.txt file:

./combine_chinese_corpus.sh

Then, split the train.txt into 8 subfiles (each for one GPU process to load):

# shuffle and split
./split.sh

2. Process SFT Dataset

2.1 Process Paper-ground Question Answering Dataset

Download the QASPER-v0.3 dataset from the path listed in ../README.md and run:

python collect.py

2.2 Process Emotional Dialogue Dataset

Simple download the scientific emotional dialogue dataset and put it under the ./data/sft/emotional folder.

2.3 Process Dolly Dataset

Download Dolly corpus by run:

cd data/sft/dolly
python download_from_hf.py

2.4 Process SciMRC Dataset

Download SciMRC dataset by the link listed in ../README, and process it by running the following command:

python collect.py

2.5 Combine Paper-ground Instruction Dataset

Combine the dolly, SciMRC and QASPER instruction dataset for supervised fine-tuning (without emotional dialogue dataset):

python combine.py