What is this repo

This repo contains the utils used by K2V to synthesize checklist.

How to use

1. After synthesizing fill-blank style QA pairs using graphgen-mask, we need to synthesize a question-specific checklist for each QA pair.

   bash data_postprocess/run.sh \
     --input_file  \
     --output_dir  \
     --model_path Qwen/Qwen2.5-72B-Instruct \
     --inference_mode offline \
     --batch_size 2000 \
     --tensor_parallel_size 8 \
     --domain EN_AGRI \

2. data_postprocess/run.sh will automatically execute a data filtering pipeline. If you want to retain the all data, you can directly run the following script.

   python data_postprocess/synthesize_checklist/synthesize_checklist.py

3. Convert the data from JSON format to Parquet format.

   python verl/convert_json_to_parquet.py

4. Verl needs a validation set to be specified to start training. Here, we choose to randomly sample from the training set as the validation set.

   python verl/get_val_dataset.py