Process dataset

Preprocessing datasets for SFT or DPO.

Quickstart

Experiment Setting

• Linux
• Python 3.10

a. Create a conda virtual environment and activate it.

conda create -n LLM_data_processing python=3.10
conda activate LLM_data_processing

b. Clone this repository.

git clone git@bitbucket.org:ibricks-rnd/llm_data_preprocess.git
cd llm_data_preprocess

c. Install requirments.

pip install -r requirements.txt

Process instruction dataset

python main.py instruction=version_4.1 main.version=ver_4.1 main.process_type=instruction

Output data format is as follow

{
  "chat_template": List[
    {
      "content": str,
      "role": str  // One of 'system', 'user', 'assistant'
    }
    ...
  ],
  "source": str,
}

In the example:

• "prompt" is a list containing dictionaries.
  • "content" is a string.
  • "role" is a string and can be one of 'system', 'user', or 'assistant'.
• "source" is a string that indicates where the data comes from."

If you want to make data for dpo, set save_data_for_dpo as true Note that, this dpo dataset has only prompt and chosen(label). you should build reject yourself.

Process DPO dataset

python main.py dpo=version_2.0_wo_chat main.version=ver_2.0_wo_chat main.process_type=dpo